Python
Monthly
Authorization bypass in praisonai-platform's workspace member removal endpoint allows any workspace member to delete any other member, including the owner. An attacker with a low-privilege member token can permanently lock the legitimate owner out of their workspace, leading to a complete denial of service and, when combined with other flawed endpoints, full workspace takeover. Patched in version 0.1.4.
Cross-workspace label tampering in praisonai-platform <=0.1.2 lets any authenticated workspace member rewrite, delete, attach, detach, and enumerate labels belonging to other tenants because five label endpoints only check workspace membership and never verify that the URL-supplied label_id or issue_id actually belongs to that workspace. Reported by the upstream MervinPraison/PraisonAI project with a fix in 0.1.4; no public exploit identified at time of analysis and the issue is not listed in CISA KEV.
Cross-workspace Insecure Direct Object Reference in praisonai-platform (<=0.1.2) allows any authenticated workspace member to read, create, and delete issue dependencies belonging to foreign workspaces by supplying arbitrary issue UUIDs in URL paths and request bodies. The dependency endpoints only enforce membership on the URL workspace_id while passing unvalidated issue and dependency identifiers to DependencyService, enabling cross-tenant linking of issues across any two workspaces in the deployment. No public exploit identified at time of analysis, but the GitHub Security Advisory GHSA-4x6r-9v57-3gqw confirms the flaw and a fixed version (0.1.4) is available.
Pre-authentication full account takeover in praisonai-platform (pip package, versions <= 0.1.2) allows remote unauthenticated attackers to forge JWTs for any user, including workspace owners and admins. The JWT signing key silently falls back to the hardcoded literal 'dev-secret-change-me' from the public GitHub source because the production-mode guard only triggers when PLATFORM_ENV is explicitly set to a non-default value, which default deployments do not do. Publicly available exploit code exists in the GHSA advisory itself, though no public exploit campaign or CISA KEV listing has been reported at time of analysis.
Vertical privilege escalation in PraisonAI Platform versions 0.1.2 and earlier allows any authenticated low-privilege workspace member to self-promote to owner and take over the workspace. The flaw stems from administrative FastAPI routes reusing a shared authorization dependency that defaults to the lowest 'member' role, so role-mutation endpoints never enforce admin/owner checks. A working PoC is published in the GHSA-h37g-4h4p-9x97 advisory, though no public exploit identified at time of analysis in mass-exploitation form, and the issue is not currently in CISA KEV.
Cross-workspace object access in PraisonAI Platform (pip package praisonai-platform <= 0.1.2) allows any authenticated workspace member to read, modify, and delete agents, projects, issues, and comments belonging to other workspaces by supplying the victim object's global UUID through their own workspace-scoped URL. The flaw stems from route-layer membership checks not being bound to service-layer object lookups, breaking tenant isolation in multi-tenant deployments. A working PoC is published in the GHSA advisory, though there is no public exploit identified at time of analysis beyond the advisory's own demonstration.
Cross-tenant IDOR and role-enforcement failure in PraisonAI Platform (praisonai-platform <= 0.1.2) let any registered user read, modify, or delete agents, issues, projects, labels, comments, and dependencies belonging to any other workspace, and let a basic member promote itself to admin/owner, evict the owner, and delete workspaces. The FastAPI `require_workspace_member` dependency validates only the URL-prefix workspace while service lookups fetch inner resources by primary key alone, so an attacker supplies their own workspace_id in the prefix and a victim's resource_id in the path. A detailed working proof-of-concept (curl-based, eight steps) is published in the vendor advisory; there is no evidence of active in-the-wild exploitation, and EPSS is low at 0.04%.
Cross-workspace IDOR and member-to-owner privilege escalation in PraisonAI Platform API (pip package praisonai-platform <= 0.1.2) lets any authenticated user read, modify, and delete issues and projects belonging to other tenants, and lets any workspace member promote themselves to owner and evict the legitimate owner. The two flaws chain together: a single member-level invite to any workspace becomes platform-wide data access plus full takeover of any workspace the attacker is added to. No public exploit identified at time of analysis beyond the detailed PoC published in the GHSA advisory, and the issue is not in CISA KEV.
Arbitrary file write in PraisonAI (pip package, versions <= 4.6.39) lets a remote attacker plant hidden metadata (output_file/output_content/save_output) in a webpage so that when a victim's PraisonAI agent crawls and analyzes it, the agent autonomously invokes write_file and drops attacker-controlled content to any absolute path. The flaw stems from write_file skipping path validation whenever workspace is None, which is the production default. Publicly available exploit code exists (a working PoC ships in the advisory); it is not listed in CISA KEV and EPSS is low (0.05%, 17th percentile), indicating no evidence of widespread active exploitation.
Remote code execution in PraisonAI's first-party A2A server example (pip package <= 4.6.39) allows unauthenticated network attackers to execute arbitrary Python on the server process by sending a JSON-RPC message/send request to /a2a that drives the LLM-backed agent into invoking an eval()-based calculate tool. The chain combines three first-party defaults - no auth_token, 0.0.0.0 bind, and an eval()-backed example tool - and was confirmed end-to-end against a real Gemini model with a canary marker file write. No public exploit identified at time of analysis beyond the proof-of-concept in the advisory itself, and the issue is not in CISA KEV.
Unauthenticated arbitrary file read in PraisonAI's MCP server (pip package praisonai, versions <= 4.6.39) lets a remote caller retrieve the full contents of any file the host user can read via the praisonai.workflow.show, workflow.validate, and deploy.validate tools. It is an incomplete-fix regression of CVE-2026-44336: the earlier patch added a path-containment helper to the rules.* handlers but left workflow.show and two adjacent handlers, plus the underlying kwargs dispatcher, unguarded. Publicly available exploit code exists (full PoC with server-start and curl attack scripts in the advisory); EPSS is very low at 0.07% (23rd percentile) and it is not listed in CISA KEV.
Remote code execution in PraisonAI praisonaiagents <=1.6.39 and PraisonAI <=4.6.39 allows authenticated attackers to fully escape the execute_code() subprocess sandbox by leveraging print.__self__ to reach the real builtins module and reconstructing __import__ at runtime. The flaw defeats prior patches for CVE-2026-39888, CVE-2026-34938, and CVE-2026-40158, enabling arbitrary OS command execution on the host wherever agent input can be influenced via prompt injection or direct code submission. Publicly available exploit code exists in the GitHub Security Advisory (GHSA-4mr5-g6f9-cfrh), and EPSS/KEV signals are not yet published for this newly disclosed novel bypass.
Authentication bypass in PraisonAI versions <= 4.6.39 allows remote unauthenticated attackers to invoke arbitrary LLM orchestration via the Flask API server generated by the documented `praisonai deploy --type api` quickstart. The generator defaults `auth_enabled=False`, causing `check_auth()` to short-circuit to `True` and accept any request to `/chat` and `/agents`, exposing the operator's LLM API keys and any agent-attached tools (python_repl, bash, file I/O, HTTP). Publicly available exploit code exists in the form of a working PoC, and CVSS scores this 9.8 critical.
Unauthenticated remote agent control in PraisonAI's call server (versions <= 4.6.39) allows any network-reachable client to enumerate, inspect, invoke, and unregister registered agents when the operator launches `praisonai-call` without setting the `CALL_SERVER_TOKEN` environment variable. The flaw stems from a fail-open authentication dependency combined with the server binding to 0.0.0.0 by default, and a working proof-of-concept is published in the GHSA advisory demonstrating end-to-end exploitation against a default deployment.
Remote code execution in PraisonAI versions 2.0.0 through 4.6.39 allows attackers to execute arbitrary Python code via two unguarded spec.loader.exec_module call sites in agents_generator.py. The flaw is a sibling of CVE-2026-44334: the v4.6.32 chokepoint refactor added a PRAISONAI_ALLOW_LOCAL_TOOLS env-var gate to tool_override.py but missed the load_tools_from_module and load_tools_from_module_class sinks, which load arbitrary module paths from YAML agent configuration without validation. Publicly available exploit code exists (working PoC published with the advisory) and a fixed release is available in 4.6.40.
Zip slip path traversal in Gotenberg through version 8.32.0 allows remote unauthenticated attackers to plant files outside the extraction directory on Windows hosts that unzip multi-output API responses. Because Gotenberg runs on Linux containers, its filepath.Base sanitisation never strips Windows-style backslashes from uploaded multipart filenames, so a crafted name like '..\..\..\Windows\System32\evil.pdf' is preserved verbatim as a zip entry name and honoured by Windows extractors (7-Zip, WinRAR, .NET ZipFile, Explorer). A working publicly available exploit code exists in the GHSA advisory; the issue is not present in CISA KEV and no EPSS score was provided.
DNS zone file injection in Froxlor (versions before 2.3.7) allows authenticated users with DNS management permissions to inject arbitrary records into bind9 zone files through incomplete validation of LOC, RP, SSHFP, and TLSA record types. This is the second attempt at fixing the issue (originally tracked as CVE-2026-30932) - the LOC regex still matches newlines via \s+, TLSA matchingType=0 accepts unbounded hex payloads, and validators return raw input without zone-file escaping. Publicly available exploit code exists demonstrating both pre-fix injection and post-fix bypasses, though no public exploit identified in active campaigns at time of analysis.
Command injection in Dulwich (pure-Python Git implementation) versions >= 0.24.0 and < 1.2.5 allows remote attackers to execute arbitrary OS commands when a victim merges an attacker-controlled branch and has a custom merge driver configured that references the %P placeholder. The ProcessMergeDriver passes attacker-controlled file paths from the git tree into subprocess.run with shell=True, so a path like 'x; touch /tmp/pwned #' is interpreted as a shell metacharacter sequence. Publicly available exploit code exists (working POC in the GitHub Security Advisory GHSA-9277-mp7x-85jf); no public exploit in CISA KEV at time of analysis.
Server-side template injection in the compliance-trestle `trestle author jinja` command enables arbitrary command execution when operators process attacker-controlled OSCAL data (SSP documents or Lookup Tables). Because the renderer recursively re-evaluates already-rendered output through a non-sandboxed Jinja2 Environment, malicious Jinja expressions placed in data fields like a system title are executed in a second pass even when the template itself is trusted and static. A proof-of-concept is published in the GHSA advisory; no public exploit identified at time of analysis as actively used in the wild, and the issue is not on CISA KEV.
Arbitrary file write in compliance-trestle's `trestle author jinja` command allows a local user supplying a crafted `-o/--output` argument to write files anywhere the invoking user can write, due to missing validation of `../`, `..\`, and absolute paths. Affected versions are <= 3.12.1 and >= 4.0.0, < 4.0.3, with fixes in 3.12.2 and 4.0.3. No public exploit identified at time of analysis, though the GitHub Security Advisory (GHSA-4q5v-7g7x-j79w) includes a full reproducer; CVSS 8.4 reflects high impact on confidentiality, integrity, and availability.
Cross-tenant data exposure in OpenReplay self-hosted session replay suite (versions prior to 1.26.0) allows an attacker holding any valid API key for their own tenant to enumerate sessions and retrieve sensitive session event data belonging to other tenants. The flaw stems from app_apikey routes in the Python API that validate the API key and the existence of a projectKey independently, but never confirm the two belong to the same tenant. No public exploit identified at time of analysis, though the trivial nature of the abuse (substituting a browser-visible projectKey) makes weaponization straightforward.
Authentication bypass in PyJWT versions prior to 2.13.0 allows remote attackers to forge valid JSON Web Tokens by exploiting an algorithm confusion flaw where the library fails to validate that a JSON Web Key intended for asymmetric verification is not reused as an HMAC shared secret. An attacker who knows the issuer's public key (typically distributed openly via JWKS endpoints) can sign HMAC-algorithm tokens with that public key and have them accepted as legitimate. No public exploit identified at time of analysis, though the underlying algorithm-confusion class is a well-documented JWT attack pattern.
Authorization bypass in OpenStack Keystone before 29.0.2 lets any authenticated user override trusted RBAC policy targets by injecting attributes like user_id or project_id into the JSON request body. The enforce_call routine unconditionally merges the raw request body over database-derived target data, so low-privileged users can perform operations on resources owned by other users or projects. No public exploit is identified at time of analysis and EPSS exploitation probability is very low (0.03%), but the flaw is trivially exploitable by any account and a vendor patch is available.
Arbitrary file write with attacker-controlled content in IBM compliance-trestle (pip package) versions up to 4.0.2 and before 3.12.2 allows a network-positioned attacker with low privilege to escape the library's cache directory by embedding path traversal sequences in OSCAL profile import URLs. The HTTPSFetcher and SFTPFetcher components in trestle/core/remote/cache.py construct local cache paths directly from URL path components without sanitizing `../` sequences, permitting writes to arbitrary filesystem locations such as /etc/cron.d or /root/.ssh/authorized_keys. Publicly available exploit code (PoC) exists in the GitHub Security Advisory GHSA-g3vg-vx23-3858; no active exploitation has been confirmed by CISA KEV, and EPSS is very low at 0.05% (15th percentile).
Remote code execution in Yamcs (the open-source mission control framework, yamcs-core) before 5.12.7 lets an authenticated operator holding the ChangeMissionDatabase privilege overwrite a Python (Jython) algorithm via the Mission Database REST API and run arbitrary OS commands on the host. The Jython script engine is invoked without a sandbox, so injected algorithm text can import java.lang.Runtime and shell out. Publicly available exploit code exists (a full PoC is published in the GitHub Security Advisory), but the issue is not listed in CISA KEV and no public in-the-wild exploitation is identified.
Remote code execution in the Yamcs mission control framework (org.yamcs:yamcs-core, releases 4.7.3 through 5.12.6) lets a caller of the algorithm-override endpoint run arbitrary Java/OS code on the ground server. The Nashorn JavaScript engine that evaluates user-supplied algorithm text is created without a ClassFilter, so payloads can reach any Java class (e.g. java.lang.Runtime) and execute commands as the Yamcs process user; because the default install (no security.yaml) gives the built-in guest user superuser=true, the endpoint is reachable by an unauthenticated network attacker. A detailed working exploit is published in the GitHub Security Advisory (publicly available exploit code exists); the issue is not listed in CISA KEV and no EPSS score was provided in the input.
Remote code execution in Langroid before 0.63.0 arises because its SQLChatAgent executes SQL text generated by an LLM, and that LLM is steerable through prompt injection — including indirect injection via data returned from the database into the model's context. When the agent connects with a database role holding code-execution or filesystem privileges, an attacker who shapes the agent's input can drive emission of dialect-specific primitives like PostgreSQL's COPY ... FROM PROGRAM to run OS commands on the database host. A full working proof-of-concept (Base64-smuggled COPY FROM PROGRAM running 'id') is published in the GitHub advisory; there is no entry in CISA KEV, so this reflects publicly available exploit code rather than confirmed active exploitation.
Unauthenticated remote code execution affects Pi.Alert, an open-source WiFi/LAN intruder detector with web-based service monitoring, in all versions prior to the 2026-05-07 release. The web configuration editor writes attacker-controlled content into pialert.conf, which the background scan daemon subsequently evaluates with Python's exec(), so injected statements run with the daemon's privileges. Because the product ships with web protection disabled by default, an attacker reaching the web interface needs no credentials, yielding a CVSS 9.8 critical flaw; no public exploit identified at time of analysis.
Unauthenticated remote code execution affects Pi.Alert, a Python-based Wi-Fi/LAN intruder detector, in all releases prior to the 2026-05-07 fix. The web UI's SaveConfigFile() endpoint writes attacker-supplied numeric configuration values such as SMTP_PORT into pialert.conf with no validation, and because that file is reloaded via Python's exec() by a background cron job every 3-5 minutes, injected Python executes at the OS level. On default installations (PIALERT_WEB_PROTECTION = False) no credentials are required, matching the CVSS 9.8 network/no-privilege rating; there is no public exploit identified at time of analysis and the CVE is not in CISA KEV, but trivial complexity and full CIA impact make it a high-priority patch.
Stored cross-site scripting in the RELATE web courseware lets any enrolled student inject JavaScript that executes in an administrator's authenticated browser session, enabling full admin account takeover. The payload is planted via the freely editable first_name/last_name fields on the /profile/ page and fires when an admin opens the Participation list in the Django admin panel. No public exploit has been identified, but the root cause is confirmed in source and fixed upstream; with a CVSS of 8.7 and a scope-changing impact, this is a high-severity privilege-escalation issue.
Unauthorized file disclosure in Taipy 4.1.1 lets remote unauthenticated attackers read files outside an extension library's intended directory through the GUI ElementLibrary.get_resource() resource handler. The containment check used str.startswith() without a trailing separator, so a crafted request with traversal segments can resolve into a prefix-matching sibling directory on disk while still passing the flawed check. Impact is confined to confidentiality (file read), with no public exploit identified at time of analysis and no CISA KEV listing.
Authentication bypass in MaxKB (1Panel-dev) versions prior to 2.9.0 allows remote unauthenticated attackers to invoke webhook trigger endpoints and execute their bound tasks. The flaw stems from the WebhookAuth class unconditionally returning a successful authentication tuple, which Django REST Framework interprets as a valid identity, combined with no backend enforcement of per-trigger token requirements. No public exploit identified at time of analysis, but the trivial nature of the bypass and open-source visibility of the patch make exploitation straightforward for any attacker who can enumerate or guess trigger IDs.
Command injection in Vowpal Wabbit's GitHub Actions CI workflow allows an
Remote code execution in HuggingFace Transformers prior to 5.3.0 allows attackers to achieve arbitrary code execution on a victim's machine by publishing a malicious model whose config.json sets the `_attn_implementation_internal` field to an attacker-controlled Hub repository. When the victim calls the standard `AutoModelForCausalLM.from_pretrained()` API, the library silently downloads and executes Python kernels from that repository with the victim's privileges, bypassing the `trust_remote_code` safety gate. No public exploit is identified at time of analysis (EPSS 0.03%, SSVC exploitation: none), but the technical impact is total and the attack uses the documented, default usage pattern.
Arbitrary code execution in Docker Desktop's Model Runner on macOS allows any container on the Docker network to achieve RCE on the host by tricking the MLX inference backend into loading a Python file from an attacker-controlled OCI model registry. The MLX-LM library imports the file referenced by config.json's model_file field via importlib without any trust_remote_code gate, and the backend runs unsandboxed as the Docker Desktop user. Patched in Docker Desktop 4.71.0; no public exploit identified at time of analysis and EPSS is very low (0.01%), but the SSVC technical impact is rated total.
Arbitrary code execution in Docker Desktop's vllm-metal inference backend on macOS allows any container on the Docker network to trigger host-level RCE by pulling a malicious model from an OCI registry and requesting inference. The Docker Model Runner unconditionally sets trust_remote_code=True and runs without sandboxing, so AutoTokenizer.from_pretrained() loads attacker-controlled Python from the model and executes it as the Docker Desktop user. No public exploit identified at time of analysis; EPSS sits at 0.01% and SSVC marks exploitation as 'none' despite total technical impact.
Unauthenticated SQL injection in YesWiki's Bazar form-import path allows any remote visitor to inject arbitrary SQL into an INSERT statement and exfiltrate the entire database, including yeswiki_users.password hashes. Affects YesWiki 4.6.1, 4.6.2, and the doryphore-dev branch prior to 4.6.4. Publicly available exploit code exists (a working Python PoC is published in the GHSA advisory), though no public exploit identified in CISA KEV at time of analysis.
Unauthenticated cross-origin MCP tool invocation in Network-AI v5.4.4 allows a remote attacker to lure a victim to a malicious web page that silently invokes any of the 22 exposed MCP tools (including config_set, agent_spawn, blackboard_write, and token_create/revoke) against the victim's locally running MCP SSE server. The vulnerability stems from an empty default secret combined with a wildcard CORS policy, and publicly available exploit code exists in the GHSA advisory demonstrating end-to-end exploitation. No CISA KEV listing yet and EPSS data was not provided, but the published PoC and trivial attack mechanics make this a meaningful risk for any user running the default Docker deployment.
Arbitrary file write on the host in Boxlite sandbox service versions prior to 0.9.0 allows attackers to escape the OCI image extraction root via crafted symlink entries in layer tarballs, enabling remote code execution on the host (typically as root). Exploitation requires a user to pull and load a malicious OCI image distributed through registries such as DockerHub. Publicly available exploit code exists (vendor-published PoC); no public exploit identified in CISA KEV at time of analysis.
Sandbox escape in Boxlite versions prior to 0.9.0 lets untrusted code running inside the lightweight VM remount host-shared virtiofs directories from read-only to read-write, enabling arbitrary writes to host files that operators believed were protected. Because the container is granted all 41 Linux capabilities (including CAP_SYS_ADMIN), a trivial 'mount -o remount,rw' bypasses the client-side MS_RDONLY enforcement, and in AI-agent deployments this leads to host code execution by tampering with mounted code, virtualenvs, or credentials. Publicly available exploit code exists (working PoC published in the GHSA advisory) and the issue carries a CVSS 10.0 with scope change; no public exploit identified at time of analysis in CISA KEV.
Unsafe default code execution in InternLM LMDeploy (<=0.12.3) lets a malicious Hugging Face model repository run arbitrary Python on the host whenever a user loads it through any LMDeploy CLI (serve, calibrate, gptq, awq). The library hardcodes transformers.AutoConfig.from_pretrained(..., trust_remote_code=True) in get_model_arch and related helpers with no flag, env var, or warning to opt out, overriding HF Transformers' default-secure stance. No public exploit identified at time of analysis, and exploitation requires the user to load an untrusted repo, so risk is hardening-level rather than network-reachable RCE.
Arbitrary code execution in InternLM lmdeploy <= 0.12.3 occurs because trust_remote_code=True is hardcoded across HuggingFace model-loading call sites in lmdeploy/archs.py and lmdeploy/utils.py. An attacker who can influence the model_path passed to an lmdeploy serving process can point it at a malicious HuggingFace repository, causing Transformers to download and execute attacker-controlled Python code with the privileges of the serving daemon. Publicly available exploit code exists in the GHSA advisory, and an upstream fix has been merged via PR #4511 (fixed in 0.13.0).
Insecure deserialization in Apache Fory's PyFory (Python) library allows remote attackers to bypass DeserializationPolicy validation hooks via the ReduceSerializer, letting untrusted classes, functions, or module attributes be restored despite policy restrictions. All PyFory releases before 1.0.0 are affected when running Python-native mode with strict mode disabled, enabling attacker-controlled data to instantiate unsafe objects and achieve code execution. There is no public exploit identified at time of analysis and EPSS probability is very low (0.04%), but CVSS is 9.8 and SSVC rates technical impact as total with automatable exploitation.
Remote code execution in Hugging Face diffusers (Python package, versions < 0.38.0) is achievable via a TOCTOU race between two sequential Hub downloads inside DiffusionPipeline.from_pretrained, letting a malicious repo owner bypass the trust_remote_code guard and silently execute arbitrary Python during model loading. Exploitation requires user interaction (loading a malicious repo without pinning a revision) and high attack complexity due to a sub-second race window, but no public exploit beyond the reporter's PoC is identified at time of analysis. Affected users running diffusers <0.38.0 should upgrade to 0.38.0 where the issue is fixed.
Unauthenticated remote code execution in 9router (npm package) versions 0.4.30 through 0.4.36 allows network-adjacent attackers to execute arbitrary OS commands by chaining two unprotected API endpoints. The Next.js authentication middleware in src/proxy.js uses a narrow route allowlist that excludes /api/cli-tools/* and /api/mcp/*, letting an attacker register an arbitrary command via POST /api/cli-tools/cowork-settings and then trigger spawn() via GET /api/mcp/[plugin]/sse. Publicly available exploit code exists (PoC published with the GHSA advisory), with CVSS 10.0 reflecting maximum severity across confidentiality, integrity, and availability.
Arbitrary file write via path traversal in Mailpit's `dump --http` subcommand (versions < 1.30.0) allows any HTTP server impersonating a Mailpit instance to write attacker-controlled bytes to arbitrary paths outside the intended output directory. The attacker controls both the file path (via the message ID field in the JSON response) and the file contents (via the raw message body endpoint), enabling writes anywhere the dumping user has write permission - including cron jobs, shell startup files, and CI artifact directories. Publicly available exploit code exists (Python PoC published in GHSA-qx5x-85p8-vg4j); no confirmed active exploitation at time of analysis.
Server-side request forgery in the zrok Python SDK ProxyShare (pip/zrok 0.4.47 through 1.1.11) lets any viewer of a share coerce the proxy host into requesting arbitrary internal or loopback services and returns the response to the attacker. Because the Flask handler forwards the client-supplied path through urllib.parse.urljoin, an absolute URL in the request path overrides the share owner's fixed target host. No public exploit has been identified at time of analysis and EPSS is low (0.06%), but the CVSS 4.0 base score is 9.9 given unauthenticated network reach.
Information disclosure in Algernon web server versions 1.17.6 and earlier allows unauthenticated remote attackers to retrieve full server-side source code, including embedded secrets, by triggering runtime errors in Lua, Pongo2, Amber, or HTML template handlers. When Algernon is started with a single file path (e.g. `algernon page.po2`), single-file mode unconditionally forces debug mode on, activating the PrettyError renderer which returns absolute file paths and complete file contents in HTTP 200 responses. Crucially, the `--prod` hardening flag does not block this behavior for non-`.lua` extensions, and publicly available exploit code exists in the GHSA advisory.
Remote code execution in APScheduler (all versions through 3.10.x and 4.0.0a5) is achievable when applications deserialize attacker-controlled data via the bundled JSONSerializer or CBORSerializer. The unmarshal_object routine dynamically imports modules and invokes __setstate__ on arbitrary classes, letting an attacker pivot an untrusted payload into code execution; publicly available exploit code exists, though EPSS remains low at 0.06% (19th percentile).
Local privilege escalation to arbitrary code execution in MLflow versions prior to 3.11.0 stems from insecure temporary directory permissions (0o777 and 0o770) created by NFS and model-download helpers. Any local user sharing the filesystem - particularly on Databricks where NFS is enabled by default - can overwrite cloudpickle-serialized model artifacts and gain code execution when another user's process deserializes them via cloudpickle.load(). No public exploit is identified at time of analysis, and the issue is a continuation of CVE-2025-10279 which was only partially fixed.
Local file disclosure in NiceGUI versions <= 3.11.1 allows remote unauthenticated attackers to read arbitrary files accessible to the server process when applications pass user-controlled content to ui.restructured_text(). The flaw stems from Docutils being invoked without disabling file-insertion directives (include, csv-table :file:, raw :file:), enabling exfiltration of secrets, credentials, and source code. No public exploit identified at time of analysis, but the vendor advisory provides full directive-level proof patterns.
Remote denial-of-service in OpenTelemetry eBPF Instrumentation (OBI) versions 0.7.0 through 0.8.x allows unauthenticated attackers to crash the privileged instrumentation process by sending a crafted memcached storage command with an oversized `<bytes>` field. The integer overflow in the memcached text protocol parser produces a negative payload length that triggers a Go runtime panic in LargeBufferReader.Peek, halting telemetry collection until OBI is restarted. Publicly available exploit code exists in the GHSA-43g7-cwr8-q3jh advisory, but there is no public exploit identified beyond the PoC and the vulnerability is not listed in CISA KEV.
Remote code execution in the amazon-redshift-python-driver (versions prior to 2.1.14) allows a malicious or compromised Redshift server, or a man-in-the-middle attacker positioned on the network path, to execute arbitrary Python code on any client that connects. The root cause is unsafe use of Python's eval() against untrusted server-supplied data inside the vector_in() function. No public exploit identified at time of analysis, but the CVSS 4.0 base score of 9.3 and PR:N/UI:N vector make this a high-priority client-side supply-chain-style risk.
Denial of service in OpenTelemetry eBPF Instrumentation (OBI) versions prior to 0.9.0 allows remote attackers to crash the telemetry agent by sending a malformed Postgres BIND frame with an empty or unterminated portal name payload to any monitored service. The defect lives in OBI's passive Postgres protocol parser, where missing NUL-terminator validation causes a Go slice-bounds panic, halting telemetry collection on the affected node. Publicly available exploit code exists in the GHSA-pgvv-q3wf-mm9m advisory, though the issue is not listed in CISA KEV and EPSS data was not provided.
Remote code execution in ChromaDB Python (version 1.0.0 and later) allows unauthenticated attackers to execute arbitrary code on the server by submitting a malicious model repository with trust_remote_code enabled via the /api/v2/tenants/{tenant}/databases/{db}/collections endpoint. The flaw carries a maximum CVSS 4.0 score of 10.0 and was disclosed publicly by HiddenLayer; no public exploit identified at time of analysis, though detailed research has been published.
Unauthenticated remote code execution in SGLang (the LLM/multimodal generation serving runtime) affecting version 5.10 arises when the non-default `--enable-custom-logit-processor` flag is set, allowing attacker-supplied Python objects to be deserialized via `dill.loads()` and execute arbitrary code on the inference host. No CISA KEV listing exists and SSVC records exploitation as 'none', but a public technical write-up (antiproof.ai, 'Three RCEs in SGLang') details the flaw and SSVC marks it automatable with total technical impact. EPSS is modest at 0.32% (55th percentile), consistent with a serious-but-conditional (feature-gated) issue rather than mass exploitation.
Budibase's REST datasource integration before version 3.38.1 bypasses IP blacklist security controls through HTTP redirect following. Authenticated Builder-level users can exploit this to access cloud metadata services and internal databases by redirecting requests through attacker-controlled servers, potentially stealing AWS/GCP/Azure credentials. This vulnerability class was previously fixed in automation steps but the REST integration was overlooked, creating an inconsistent security posture.
Path traversal in Pipecat's development runner allows unauthenticated remote attackers to read arbitrary files when the --folder flag is used. The /files/{filename:path} endpoint fails to validate paths containing %2F-encoded directory separators, bypassing Starlette's URL normalization. Fixed in version 1.2.0 with no public exploit identified at time of analysis.
Authentication bypass in MLflow 3.9.0 and earlier allows unauthenticated remote attackers to access protected Job API and OpenTelemetry trace ingestion endpoints when the server runs with basic-auth enabled via uvicorn/ASGI. Attackers can submit jobs, read results, cancel operations, and inject trace data without credentials. The FastAPI permission middleware incorrectly enforced authentication only on /gateway/ routes, leaving /ajax-api/3.0/jobs/* and /v1/traces unprotected due to architectural mismatch between Flask and FastAPI authentication mechanisms. Fixed in version 3.10.0 with GitHub commit bb62e77 adding proper validators for all FastAPI routes.
python-utcp CLI subprocess environment passes all process-level secrets to every tool call. When chained with CVE-2026-45369 command injection, remote authenticated attackers with low-privilege LLM tool access can exfiltrate AWS credentials, API keys, database URLs, and other environment variables in a single HTTP request. Patch available in version 1.1.2 (NVD references 1.1.3 as fixed version). GitHub security advisory confirms proof-of-concept demonstrating credential theft via env dump to attacker-controlled endpoint.
Command injection in python-utcp allows remote attackers to execute arbitrary shell commands on Unix and Windows systems when user-controlled tool arguments are processed by the CLI communication protocol module. The _substitute_utcp_args method in cli_communication_protocol.py directly embeds unsanitized user input into bash or PowerShell commands without escaping, enabling full remote code execution. Vendor-released patch available in version 1.1.2 with shell-quoting mitigation (shlex.quote on Unix, single-quoted literals on Windows). CVSS 8.3 indicates high complexity and required user interaction, but scope change enables container/sandbox escape scenarios. No public exploit code or CISA KEV listing identified at time of analysis, though detailed proof-of-concept exists in the GitHub security advisory demonstrating data exfiltration via curl.
Multiple concurrent LDAP or OAuth first-login requests on a freshly deployed Open WebUI instance can all receive administrator privileges through a TOCTOU race condition in role assignment logic. The vulnerability affects deployments using LDAP or OAuth authentication on instances with no existing users. While the regular signup handler was explicitly patched for this race condition in earlier code ('Insert with default role first to avoid TOCTOU race'), the LDAP and OAuth authentication paths were never updated with the same fix. Vendor-released patch available in version 0.9.0 (April 2026). No active exploitation confirmed (not in CISA KEV), though publicly available exploit code exists per GitHub advisory GHSA-h3ww-q6xx-w7x3. CVSS 8.1 (High) reflects network attack vector but requires high attack complexity (precise timing of concurrent requests during narrow first-deployment window).
Open WebUI versions through 0.8.11 allow authenticated users to execute arbitrary Python code in the Jupyter container by bypassing the ENABLE_CODE_EXECUTION=false configuration flag. The /api/v1/utils/code/execute endpoint fails to enforce the admin-configured feature gate (CWE-863: Incorrect Authorization), enabling any verified user to run code even when administrators believe execution is disabled. The vulnerability is confirmed by vendor POC (verified 2026-03-25) demonstrating successful code execution, file access, and SSRF to internal Docker services despite explicit admin configuration disabling the feature. Vendor-released patch available in v0.8.12 (commit 6d736d3c5) enforces the configuration check before dispatching code to Jupyter.
Broken object-level authorization in Open WebUI versions ≤0.8.12 allows any authenticated user to permanently delete files owned by other users when those files are referenced in any shared chat. The has_access_to_file() authorization function unconditionally grants access through its shared-chat branch, failing to validate both the requesting user's identity and the operation type (read vs. write). File UUIDs, which would otherwise prevent enumeration attacks, are exposed via the knowledge base API endpoint GET /api/v1/knowledge/{id}/files to any user with read access. This affects all default Docker deployments where chat sharing is enabled. Vendor-released patch available in v0.9.0 (commit 2e52ad8ff). No active exploitation confirmed (not in CISA KEV). CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:U/C:H/I:H/A:H scores 8.0, though real-world impact extends beyond confidentiality to permanent data destruction with no recovery mechanism.
URL parser mismatch in Open WebUI allows authenticated users to bypass SSRF protections and access internal network resources. The validate_url function uses Python's urlparse library to extract hostnames for validation, while the requests library handles actual HTTP requests. These libraries disagree on parsing URLs containing backslash characters (e.g., http://127.0.0.1:6666\@1.1.1.1), allowing attackers to craft URLs that pass validation as external addresses but resolve to internal hosts. Exploitation requires low-privilege authentication but no user interaction, enabling access to cloud metadata endpoints and internal services. Fixed in version 0.9.5 per GitHub advisory GHSA-8w7q-q5jp-jvgx.
Broken object-level authorization in Open WebUI allows any authenticated low-privilege user to enumerate and terminate background tasks belonging to other users via GET /api/tasks and POST /api/tasks/stop/{task_id}. Attackers can disrupt system-wide chat generation and background processing by continuously canceling active tasks across the multi-user instance. Publicly available exploit code exists. Vendor-released patch in v0.9.0 restricts global task endpoints to admin-only and introduces a scoped /api/tasks/chat/{chat_id}/stop endpoint for legitimate user-owned task termination. CVSS 7.1 (AV:N/AC:L/PR:L/UI:N) reflects network-accessible, low-complexity exploitation requiring only authenticated low-privilege access, with high availability impact and low confidentiality impact from task enumeration.
Insecure Direct Object Reference (IDOR) in Open WebUI's retrieval API allows authenticated users to bypass knowledge base access controls and directly access, modify, or delete other users' private knowledge bases by supplying the target UUID as a collection name. The authorization gap affects seven endpoints: two read endpoints (/query/doc, /query/collection) permit exfiltration of private knowledge base content, while five write endpoints (/process/text, /process/file, /process/files/batch, /process/web, /process/youtube) enable content injection, poisoning, or complete data destruction via overwrite. Affects Open WebUI <= 0.9.4; fixed in v0.9.5 via PR #22109. EPSS data not available; no confirmed active exploitation (CVSS 7.5 reflects AC:H due to UUID prerequisite, but UUIDs leak through multiple channels per researcher analysis).
Privilege escalation in Open WebUI allows authenticated users with 'write' access grants on tools to execute arbitrary server-side Python code without the required workspace.tools permission. The tool update endpoint (POST /api/v1/tools/id/{id}/update) fails to enforce the workspace.tools authorization check that gates code execution, allowing users explicitly denied code execution capabilities to bypass this security boundary. This breaks Open WebUI's documented trust model where workspace.tools permission is intentionally disabled by default and 'equivalent to giving them shell access to the server.' Exploitation achieves root code execution (PID 1) in default Docker deployments, enabling extraction of secrets (WEBUI_SECRET_KEY, API keys), database access, and filesystem read/write. Confirmed by GitHub security advisory GHSA-p4fx-23fq-jfg6. No public exploit or KEV listing at time of analysis, but detailed proof-of-concept with Burp Collaborator confirmation exists in the advisory.
Server-Side Request Forgery (SSRF) in Open WebUI versions ≤0.8.12 allows authenticated users with OAuth access to force the server to make HTTP requests to arbitrary internal resources and exfiltrate complete response data. Exploitation requires OAuth-enabled deployments with ENABLE_OAUTH_SIGNUP=true or OAUTH_UPDATE_PICTURE_ON_LOGIN=true. An attacker controls the OAuth provider's 'picture' claim URL, triggering server-side HTTP requests to cloud metadata services (AWS IMDS), localhost services (Redis, Elasticsearch), or internal network endpoints. The full response is base64-encoded and stored in the user's profile_image_url field, enabling complete data exfiltration. Fixed in version 0.9.0 per GitHub advisory GHSA-24c9-2m8q-qhmh. EPSS data not available; no CISA KEV listing indicates limited widespread exploitation, though publicly available proof-of-concept exists in the GitHub advisory.
Server-Side Request Forgery in Open WebUI's `validate_url()` function allows authenticated attackers to reach internal IPv4/IPv6 addresses, bypassing security controls via three distinct flaws: the validators library silently fails on IPv6 private-address checks (raising ValidationError which evaluates as falsy), IPv4-mapped IPv6 addresses (::ffff:10.0.0.1) evade IPv4 filtering entirely, and multiple IANA-reserved IPv4 ranges (0.0.0.0/8, 100.64.0.0/10, 192.0.0.0/24, 198.18.0.0/15, 203.0.113.0/24) remain unblocked. The vulnerability persists in the RAG web search, image editing, and other endpoints despite an earlier incomplete remediation attempt (CVE-2025-65958), enabling exfiltration of AWS IMDSv1 credentials and access to localhost-bound services. Publicly available exploit code exists (demonstrated POC in advisory), affecting Open WebUI ≤0.8.12 with fix released in version 0.9.0.
Stored cross-site scripting (XSS) in Open WebUI ≤0.9.2 allows authenticated users with default speech-to-text permissions to upload polyglot WAV+HTML files through the audio transcription endpoint, achieving code execution in victim browsers and enabling full account takeover including administrator sessions. The vulnerability chains insecure file extension handling with unrestricted Content-Type serving and non-HttpOnly JWT storage to weaponize a single-click attack. Publicly available exploit code exists with video demonstration; no active exploitation confirmed by CISA KEV at time of analysis. CVSS 8.7 (High) reflects changed scope (S:C) and user interaction requirement, but real-world risk is elevated because the vulnerable permission defaults to enabled and the attack yields immediate admin-level access in typical deployments.
Cleartext HMAC signing key exposure in Amazon SageMaker Python SDK versions <2.257.2 and <3.8.0 enables authenticated attackers with SageMaker describe API and S3 write permissions to forge model artifact integrity signatures and achieve remote code execution in inference containers. AWS released patches in v2.257.2 and v3.8.0 with security fixes addressing Triton HMAC key exposure and missing integrity checks. EPSS data not available; no CISA KEV listing or public POC identified at time of analysis, suggesting limited exploitation activity despite high CVSS score.
Authenticated server-side request forgery in ApostropheCMS allows low-privilege users to force the server to fetch arbitrary internal URLs through the rich-text widget import flow. Attackers with content editing permissions can exfiltrate internal data by crafting malicious image tags that trigger server-side fetch operations, with image-compatible responses being persisted and re-hosted by the application. Publicly available exploit code exists (full Python PoC published in GitHub advisory GHSA-pr28-mf3q-qpg6), enabling immediate weaponization. All versions through 4.29.0 are affected with no vendor-released patch identified at time of analysis, creating sustained exposure for organizations running this popular Node.js CMS.
Remote code execution in HuggingFace Diffusers library (versions < 0.38.0) allows attackers to execute arbitrary Python code when victims load malicious pipelines from Hugging Face Hub repositories. The vulnerability bypasses the trust_remote_code=True safeguard through a type coercion flaw where None values are interpolated as 'None.py' filenames. Attackers can achieve silent code execution by publishing repositories containing a malicious None.py file alongside legitimate-looking configuration, requiring only that victims call DiffusionPipeline.from_pretrained() on the attacker's repository. EPSS data not available; no public exploit identified at time of analysis. Vendor-released patch: version 0.38.0.
Privilege escalation in wger fitness manager allows gym trainers to impersonate gym managers via session-chain attack. An authenticated trainer exploits flawed session-flag logic in the trainer-login endpoint to bypass permission checks - first switching into a low-privilege user, then leveraging the inherited 'trainer.identity' session flag to hop into manager accounts. Publicly available proof-of-concept demonstrates complete takeover of gym administration with CVSS 8.1 (network-accessible, low complexity). No vendor patch confirmed at time of analysis; vulnerability actively disclosed by wger-project GitHub advisory GHSA-9qpr-vc49-hqg2. EPSS score not available, not in CISA KEV. Root cause is CWE-269 (improper privilege management) in core/views/user.py lines 169-178.
Insecure Direct Object Reference in wger fitness platform exposes any authenticated user's complete workout history, session notes, and training statistics through template routine API endpoints. Attackers with free accounts enumerate public template routine IDs and retrieve owners' private health data including workout notes, weights, repetitions, and performance metrics via /api/v2/routine/{id}/logs/ and /api/v2/routine/{id}/stats/ endpoints. Detailed proof-of-concept with Python exploit confirms trivial exploitation against wger <= 2.5.0a2. CVSS 7.5 rates this High severity, but NOTE: vector PR:N appears inconsistent with authenticated-only access described - attackers need valid credentials, suggesting actual vector should be PR:L. EPSS data not available. No CVE KEV listing or public exploit repositories identified beyond GitHub advisory disclosure. Patch status unconfirmed - GitHub advisory references fix commit but no released version number provided in available data.
Local code execution in the claude-code-cache-fix npm package (v3.5.0 and v3.5.1) lets attacker-controlled filesystem path names run arbitrary Python inside a victim's Claude Code process. The bundled tools/quota-statusline.sh interpolates Claude Code's statusline hook stdin — which reflects user-controlled paths such as cwd, workspace.current_dir, workspace.project_dir, and transcript_path — directly into a Python triple-quoted literal, so a directory name containing the byte sequence ''' closes the literal early and executes following bytes as Python at the user's privilege on every statusline redraw. A working injection payload is publicly available exploit code (published in the GHSA advisory and the T6/T7 regression tests); the issue is not listed in CISA KEV and no EPSS score was provided.
Unsafe deserialization in LangSmith SDK's prompt pull methods allows remote attackers to execute server-side request forgery (SSRF) and redirect LLM traffic to attacker-controlled infrastructure when applications pull public prompts from LangSmith Hub. The SDK deserializes untrusted prompt manifests containing serialized LangChain objects with attacker-controlled constructor arguments, including malicious base_url configurations, custom headers, and secret references. Exploitation requires user interaction (developers must call pull_prompt with a malicious owner/name identifier), but no authentication is required to publish malicious prompts to the public Hub. Vendor-released patches in Python >= 0.8.0 and JS/TS >= 0.6.0 now block public prompt pulling by default, requiring explicit opt-in via dangerously_pull_public_prompt flag. EPSS data not available; no CISA KEV listing or public exploit identified at time of analysis.
Denial-of-service via memory exhaustion affects UltraJSON (ujson) versions 5.12.0 and earlier, where the ujson.dump() function fails to release (Py_DECREF) the serialized JSON string object when the target file-like object's write() method raises an exception. Each failed write leaks the full serialized payload, so an attacker who can repeatedly trigger write failures - for example by disconnecting mid-response from a web server that streams JSON via ujson.dump() - can drive linear, unbounded memory growth. A proof-of-concept exists in the vendor advisory (GHSA-c38f-wx89-p2xg), but there is no public exploit identified as weaponized and no active exploitation; EPSS is very low at 0.04% (12th percentile).
Sandbox escape in Heym's custom Python tool executor (versions before 0.0.21) allows authenticated workflow authors to break out of the restricted execution environment using Python object-graph introspection, recover the unrestricted __import__ function, and load blocked modules such as os and subprocess. Once free of the sandbox, an attacker executes arbitrary host commands as the backend service user and reads inherited environment variables holding database credentials and encryption keys. Publicly available exploit code exists (SSVC 'poc'), though the EPSS probability is very low at 0.04% and it is not listed in CISA KEV.
Sandbox escape in OpenClaude (npm package openclaude) versions before 0.5.1 allows a prompt-injected LLM to disable host sandboxing by setting the model-controlled `dangerouslyDisableSandbox: true` flag in any Bash tool_use call, yielding full unsandboxed command execution on the host. CVSS 4.0 scores this 9.3 Critical (AV:N/AC:L/PR:N/UI:N, VC/VI/VA:H); no public exploit identified at time of analysis beyond the reporter's PoC, but the upstream fix has been merged. The flaw is especially severe because it is reachable under default settings (`allowUnsandboxedCommands` defaults to true).
Arbitrary code execution via torch-checkpoint-shrink.py script in ml-engineering project allows remote attackers to execute malicious Python code by providing crafted PyTorch checkpoint files. The vulnerability stems from insecure deserialization where torch.load() processes .pt files without the weights_only=True safeguard, enabling pickle-based arbitrary object instantiation. Despite a critical CVSS 9.8 score, EPSS probability is low (0.06%, 19th percentile) and no public exploit or active exploitation is confirmed, suggesting limited real-world targeting to date. SSVC assessment indicates total technical impact with automatable exploitation potential, making this a priority for organizations using ml-engineering scripts in production environments.
Remote code execution in Ludwig framework ≤0.10.4 allows unauthenticated network attackers to execute arbitrary code by supplying a malicious PyTorch model file to the ludwig serve endpoint. The vulnerability stems from unsafe deserialization in the model loading component, which uses torch.load() without the weights_only=True safety parameter. With CVSS 9.8 (critical network vector, no authentication required) but only 0.02% EPSS, this represents a high-severity issue in vulnerable deployments, though widespread exploitation has not been observed. No CISA KEV listing or public POC identified at time of analysis.
Arbitrary code execution in Ludwig framework ≤0.10.4 occurs when attackers supply malicious pickle files to the predict() method, which deserializes untrusted data without validation using pandas.read_pickle(). Remote unauthenticated attackers can achieve full system compromise by exploiting the automatic file format detection mechanism that processes .pkl files through Python's unsafe pickle module. EPSS score of 0.06% (19th percentile) suggests low current exploitation likelihood despite the critical CVSS 9.8 rating, though no public exploit code or active exploitation has been identified at time of analysis.
Remote code execution in Mamba language model framework (through version 2.2.6) allows unauthenticated attackers to execute arbitrary Python code by publishing malicious models on HuggingFace Hub. When victims call MambaLMHeadModel.from_pretrained() on a weaponized model repository, insecure pickle deserialization executes attacker-controlled code in the context of the victim's process. Despite the critical CVSS 9.8 score and network attack vector requiring no authentication, EPSS probability remains extremely low (0.02%, 5th percentile), suggesting limited real-world exploitation to date. No CISA KEV listing or public POC identified at time of analysis.
Remote code execution in Optimate's neural_magic_training.py script allows authenticated attackers to execute arbitrary code via malicious PyTorch model files. The vulnerability stems from unsafe deserialization when loading model state dictionaries without PyTorch's weights_only=True security flag, enabling pickle-based arbitrary object execution. With an EPSS score of 0.06% and no confirmed exploitation, this represents a moderate risk primarily in environments where users can upload or specify model files.
Arbitrary code execution in optimate's neural_magic_training.py allows remote attackers to execute Python code by supplying a malicious directory path containing a crafted module.py file. The _load_model() function directly executes file contents via Python's exec() without validation. CVSS 9.8 reflects network vector with no authentication, but EPSS score of 0.02% (5th percentile) indicates very low observed exploitation probability. No active exploitation confirmed (not in CISA KEV). Vulnerability exists in commit a6d302f912b481c94370811af6b11402f51d377f from July 2024. Affects organizations using optimate for neural network model optimization.
Remote code execution in PySyft Datasite/Server versions 0.9.5 and earlier allows unauthenticated attackers to execute arbitrary Python code on the server through the function submission mechanism. The vulnerability stems from insufficient validation and sandboxing of user-submitted Python functions decorated with @sy.syft_function(), which are executed using unsafe exec() and eval() calls after approval. With an EPSS score of 0.04% and no current KEV listing, this appears to be a high-severity vulnerability without confirmed active exploitation.
Remote code execution in Snorkel machine learning library (≤v0.10.0) occurs when users load untrusted model files via MultitaskClassifier.load(). The vulnerability exploits insecure Python object deserialization through torch.load(), allowing attackers to embed malicious code in model weight files that executes upon loading. EPSS score of 0.06% (19th percentile) suggests low observed exploitation probability in the wild, though SSVC framework indicates total technical impact once exploited. No public exploit code or active exploitation confirmed at time of analysis, but exploitation requires only that a data scientist or ML engineer load a malicious .pkl model file.
Insecure deserialization in Optimate's neural_magic_training.py script enables remote code execution when loading PyTorch model files. The _load_model() function uses torch.load() without the weights_only=True security parameter, allowing attackers with low privileges to execute arbitrary Python code by providing malicious .pt or .pth files via the --model command-line argument. EPSS indicates low exploitation probability at 0.06% with no active exploitation confirmed.
Authorization bypass in praisonai-platform's workspace member removal endpoint allows any workspace member to delete any other member, including the owner. An attacker with a low-privilege member token can permanently lock the legitimate owner out of their workspace, leading to a complete denial of service and, when combined with other flawed endpoints, full workspace takeover. Patched in version 0.1.4.
Cross-workspace label tampering in praisonai-platform <=0.1.2 lets any authenticated workspace member rewrite, delete, attach, detach, and enumerate labels belonging to other tenants because five label endpoints only check workspace membership and never verify that the URL-supplied label_id or issue_id actually belongs to that workspace. Reported by the upstream MervinPraison/PraisonAI project with a fix in 0.1.4; no public exploit identified at time of analysis and the issue is not listed in CISA KEV.
Cross-workspace Insecure Direct Object Reference in praisonai-platform (<=0.1.2) allows any authenticated workspace member to read, create, and delete issue dependencies belonging to foreign workspaces by supplying arbitrary issue UUIDs in URL paths and request bodies. The dependency endpoints only enforce membership on the URL workspace_id while passing unvalidated issue and dependency identifiers to DependencyService, enabling cross-tenant linking of issues across any two workspaces in the deployment. No public exploit identified at time of analysis, but the GitHub Security Advisory GHSA-4x6r-9v57-3gqw confirms the flaw and a fixed version (0.1.4) is available.
Pre-authentication full account takeover in praisonai-platform (pip package, versions <= 0.1.2) allows remote unauthenticated attackers to forge JWTs for any user, including workspace owners and admins. The JWT signing key silently falls back to the hardcoded literal 'dev-secret-change-me' from the public GitHub source because the production-mode guard only triggers when PLATFORM_ENV is explicitly set to a non-default value, which default deployments do not do. Publicly available exploit code exists in the GHSA advisory itself, though no public exploit campaign or CISA KEV listing has been reported at time of analysis.
Vertical privilege escalation in PraisonAI Platform versions 0.1.2 and earlier allows any authenticated low-privilege workspace member to self-promote to owner and take over the workspace. The flaw stems from administrative FastAPI routes reusing a shared authorization dependency that defaults to the lowest 'member' role, so role-mutation endpoints never enforce admin/owner checks. A working PoC is published in the GHSA-h37g-4h4p-9x97 advisory, though no public exploit identified at time of analysis in mass-exploitation form, and the issue is not currently in CISA KEV.
Cross-workspace object access in PraisonAI Platform (pip package praisonai-platform <= 0.1.2) allows any authenticated workspace member to read, modify, and delete agents, projects, issues, and comments belonging to other workspaces by supplying the victim object's global UUID through their own workspace-scoped URL. The flaw stems from route-layer membership checks not being bound to service-layer object lookups, breaking tenant isolation in multi-tenant deployments. A working PoC is published in the GHSA advisory, though there is no public exploit identified at time of analysis beyond the advisory's own demonstration.
Cross-tenant IDOR and role-enforcement failure in PraisonAI Platform (praisonai-platform <= 0.1.2) let any registered user read, modify, or delete agents, issues, projects, labels, comments, and dependencies belonging to any other workspace, and let a basic member promote itself to admin/owner, evict the owner, and delete workspaces. The FastAPI `require_workspace_member` dependency validates only the URL-prefix workspace while service lookups fetch inner resources by primary key alone, so an attacker supplies their own workspace_id in the prefix and a victim's resource_id in the path. A detailed working proof-of-concept (curl-based, eight steps) is published in the vendor advisory; there is no evidence of active in-the-wild exploitation, and EPSS is low at 0.04%.
Cross-workspace IDOR and member-to-owner privilege escalation in PraisonAI Platform API (pip package praisonai-platform <= 0.1.2) lets any authenticated user read, modify, and delete issues and projects belonging to other tenants, and lets any workspace member promote themselves to owner and evict the legitimate owner. The two flaws chain together: a single member-level invite to any workspace becomes platform-wide data access plus full takeover of any workspace the attacker is added to. No public exploit identified at time of analysis beyond the detailed PoC published in the GHSA advisory, and the issue is not in CISA KEV.
Arbitrary file write in PraisonAI (pip package, versions <= 4.6.39) lets a remote attacker plant hidden metadata (output_file/output_content/save_output) in a webpage so that when a victim's PraisonAI agent crawls and analyzes it, the agent autonomously invokes write_file and drops attacker-controlled content to any absolute path. The flaw stems from write_file skipping path validation whenever workspace is None, which is the production default. Publicly available exploit code exists (a working PoC ships in the advisory); it is not listed in CISA KEV and EPSS is low (0.05%, 17th percentile), indicating no evidence of widespread active exploitation.
Remote code execution in PraisonAI's first-party A2A server example (pip package <= 4.6.39) allows unauthenticated network attackers to execute arbitrary Python on the server process by sending a JSON-RPC message/send request to /a2a that drives the LLM-backed agent into invoking an eval()-based calculate tool. The chain combines three first-party defaults - no auth_token, 0.0.0.0 bind, and an eval()-backed example tool - and was confirmed end-to-end against a real Gemini model with a canary marker file write. No public exploit identified at time of analysis beyond the proof-of-concept in the advisory itself, and the issue is not in CISA KEV.
Unauthenticated arbitrary file read in PraisonAI's MCP server (pip package praisonai, versions <= 4.6.39) lets a remote caller retrieve the full contents of any file the host user can read via the praisonai.workflow.show, workflow.validate, and deploy.validate tools. It is an incomplete-fix regression of CVE-2026-44336: the earlier patch added a path-containment helper to the rules.* handlers but left workflow.show and two adjacent handlers, plus the underlying kwargs dispatcher, unguarded. Publicly available exploit code exists (full PoC with server-start and curl attack scripts in the advisory); EPSS is very low at 0.07% (23rd percentile) and it is not listed in CISA KEV.
Remote code execution in PraisonAI praisonaiagents <=1.6.39 and PraisonAI <=4.6.39 allows authenticated attackers to fully escape the execute_code() subprocess sandbox by leveraging print.__self__ to reach the real builtins module and reconstructing __import__ at runtime. The flaw defeats prior patches for CVE-2026-39888, CVE-2026-34938, and CVE-2026-40158, enabling arbitrary OS command execution on the host wherever agent input can be influenced via prompt injection or direct code submission. Publicly available exploit code exists in the GitHub Security Advisory (GHSA-4mr5-g6f9-cfrh), and EPSS/KEV signals are not yet published for this newly disclosed novel bypass.
Authentication bypass in PraisonAI versions <= 4.6.39 allows remote unauthenticated attackers to invoke arbitrary LLM orchestration via the Flask API server generated by the documented `praisonai deploy --type api` quickstart. The generator defaults `auth_enabled=False`, causing `check_auth()` to short-circuit to `True` and accept any request to `/chat` and `/agents`, exposing the operator's LLM API keys and any agent-attached tools (python_repl, bash, file I/O, HTTP). Publicly available exploit code exists in the form of a working PoC, and CVSS scores this 9.8 critical.
Unauthenticated remote agent control in PraisonAI's call server (versions <= 4.6.39) allows any network-reachable client to enumerate, inspect, invoke, and unregister registered agents when the operator launches `praisonai-call` without setting the `CALL_SERVER_TOKEN` environment variable. The flaw stems from a fail-open authentication dependency combined with the server binding to 0.0.0.0 by default, and a working proof-of-concept is published in the GHSA advisory demonstrating end-to-end exploitation against a default deployment.
Remote code execution in PraisonAI versions 2.0.0 through 4.6.39 allows attackers to execute arbitrary Python code via two unguarded spec.loader.exec_module call sites in agents_generator.py. The flaw is a sibling of CVE-2026-44334: the v4.6.32 chokepoint refactor added a PRAISONAI_ALLOW_LOCAL_TOOLS env-var gate to tool_override.py but missed the load_tools_from_module and load_tools_from_module_class sinks, which load arbitrary module paths from YAML agent configuration without validation. Publicly available exploit code exists (working PoC published with the advisory) and a fixed release is available in 4.6.40.
Zip slip path traversal in Gotenberg through version 8.32.0 allows remote unauthenticated attackers to plant files outside the extraction directory on Windows hosts that unzip multi-output API responses. Because Gotenberg runs on Linux containers, its filepath.Base sanitisation never strips Windows-style backslashes from uploaded multipart filenames, so a crafted name like '..\..\..\Windows\System32\evil.pdf' is preserved verbatim as a zip entry name and honoured by Windows extractors (7-Zip, WinRAR, .NET ZipFile, Explorer). A working publicly available exploit code exists in the GHSA advisory; the issue is not present in CISA KEV and no EPSS score was provided.
DNS zone file injection in Froxlor (versions before 2.3.7) allows authenticated users with DNS management permissions to inject arbitrary records into bind9 zone files through incomplete validation of LOC, RP, SSHFP, and TLSA record types. This is the second attempt at fixing the issue (originally tracked as CVE-2026-30932) - the LOC regex still matches newlines via \s+, TLSA matchingType=0 accepts unbounded hex payloads, and validators return raw input without zone-file escaping. Publicly available exploit code exists demonstrating both pre-fix injection and post-fix bypasses, though no public exploit identified in active campaigns at time of analysis.
Command injection in Dulwich (pure-Python Git implementation) versions >= 0.24.0 and < 1.2.5 allows remote attackers to execute arbitrary OS commands when a victim merges an attacker-controlled branch and has a custom merge driver configured that references the %P placeholder. The ProcessMergeDriver passes attacker-controlled file paths from the git tree into subprocess.run with shell=True, so a path like 'x; touch /tmp/pwned #' is interpreted as a shell metacharacter sequence. Publicly available exploit code exists (working POC in the GitHub Security Advisory GHSA-9277-mp7x-85jf); no public exploit in CISA KEV at time of analysis.
Server-side template injection in the compliance-trestle `trestle author jinja` command enables arbitrary command execution when operators process attacker-controlled OSCAL data (SSP documents or Lookup Tables). Because the renderer recursively re-evaluates already-rendered output through a non-sandboxed Jinja2 Environment, malicious Jinja expressions placed in data fields like a system title are executed in a second pass even when the template itself is trusted and static. A proof-of-concept is published in the GHSA advisory; no public exploit identified at time of analysis as actively used in the wild, and the issue is not on CISA KEV.
Arbitrary file write in compliance-trestle's `trestle author jinja` command allows a local user supplying a crafted `-o/--output` argument to write files anywhere the invoking user can write, due to missing validation of `../`, `..\`, and absolute paths. Affected versions are <= 3.12.1 and >= 4.0.0, < 4.0.3, with fixes in 3.12.2 and 4.0.3. No public exploit identified at time of analysis, though the GitHub Security Advisory (GHSA-4q5v-7g7x-j79w) includes a full reproducer; CVSS 8.4 reflects high impact on confidentiality, integrity, and availability.
Cross-tenant data exposure in OpenReplay self-hosted session replay suite (versions prior to 1.26.0) allows an attacker holding any valid API key for their own tenant to enumerate sessions and retrieve sensitive session event data belonging to other tenants. The flaw stems from app_apikey routes in the Python API that validate the API key and the existence of a projectKey independently, but never confirm the two belong to the same tenant. No public exploit identified at time of analysis, though the trivial nature of the abuse (substituting a browser-visible projectKey) makes weaponization straightforward.
Authentication bypass in PyJWT versions prior to 2.13.0 allows remote attackers to forge valid JSON Web Tokens by exploiting an algorithm confusion flaw where the library fails to validate that a JSON Web Key intended for asymmetric verification is not reused as an HMAC shared secret. An attacker who knows the issuer's public key (typically distributed openly via JWKS endpoints) can sign HMAC-algorithm tokens with that public key and have them accepted as legitimate. No public exploit identified at time of analysis, though the underlying algorithm-confusion class is a well-documented JWT attack pattern.
Authorization bypass in OpenStack Keystone before 29.0.2 lets any authenticated user override trusted RBAC policy targets by injecting attributes like user_id or project_id into the JSON request body. The enforce_call routine unconditionally merges the raw request body over database-derived target data, so low-privileged users can perform operations on resources owned by other users or projects. No public exploit is identified at time of analysis and EPSS exploitation probability is very low (0.03%), but the flaw is trivially exploitable by any account and a vendor patch is available.
Arbitrary file write with attacker-controlled content in IBM compliance-trestle (pip package) versions up to 4.0.2 and before 3.12.2 allows a network-positioned attacker with low privilege to escape the library's cache directory by embedding path traversal sequences in OSCAL profile import URLs. The HTTPSFetcher and SFTPFetcher components in trestle/core/remote/cache.py construct local cache paths directly from URL path components without sanitizing `../` sequences, permitting writes to arbitrary filesystem locations such as /etc/cron.d or /root/.ssh/authorized_keys. Publicly available exploit code (PoC) exists in the GitHub Security Advisory GHSA-g3vg-vx23-3858; no active exploitation has been confirmed by CISA KEV, and EPSS is very low at 0.05% (15th percentile).
Remote code execution in Yamcs (the open-source mission control framework, yamcs-core) before 5.12.7 lets an authenticated operator holding the ChangeMissionDatabase privilege overwrite a Python (Jython) algorithm via the Mission Database REST API and run arbitrary OS commands on the host. The Jython script engine is invoked without a sandbox, so injected algorithm text can import java.lang.Runtime and shell out. Publicly available exploit code exists (a full PoC is published in the GitHub Security Advisory), but the issue is not listed in CISA KEV and no public in-the-wild exploitation is identified.
Remote code execution in the Yamcs mission control framework (org.yamcs:yamcs-core, releases 4.7.3 through 5.12.6) lets a caller of the algorithm-override endpoint run arbitrary Java/OS code on the ground server. The Nashorn JavaScript engine that evaluates user-supplied algorithm text is created without a ClassFilter, so payloads can reach any Java class (e.g. java.lang.Runtime) and execute commands as the Yamcs process user; because the default install (no security.yaml) gives the built-in guest user superuser=true, the endpoint is reachable by an unauthenticated network attacker. A detailed working exploit is published in the GitHub Security Advisory (publicly available exploit code exists); the issue is not listed in CISA KEV and no EPSS score was provided in the input.
Remote code execution in Langroid before 0.63.0 arises because its SQLChatAgent executes SQL text generated by an LLM, and that LLM is steerable through prompt injection — including indirect injection via data returned from the database into the model's context. When the agent connects with a database role holding code-execution or filesystem privileges, an attacker who shapes the agent's input can drive emission of dialect-specific primitives like PostgreSQL's COPY ... FROM PROGRAM to run OS commands on the database host. A full working proof-of-concept (Base64-smuggled COPY FROM PROGRAM running 'id') is published in the GitHub advisory; there is no entry in CISA KEV, so this reflects publicly available exploit code rather than confirmed active exploitation.
Unauthenticated remote code execution affects Pi.Alert, an open-source WiFi/LAN intruder detector with web-based service monitoring, in all versions prior to the 2026-05-07 release. The web configuration editor writes attacker-controlled content into pialert.conf, which the background scan daemon subsequently evaluates with Python's exec(), so injected statements run with the daemon's privileges. Because the product ships with web protection disabled by default, an attacker reaching the web interface needs no credentials, yielding a CVSS 9.8 critical flaw; no public exploit identified at time of analysis.
Unauthenticated remote code execution affects Pi.Alert, a Python-based Wi-Fi/LAN intruder detector, in all releases prior to the 2026-05-07 fix. The web UI's SaveConfigFile() endpoint writes attacker-supplied numeric configuration values such as SMTP_PORT into pialert.conf with no validation, and because that file is reloaded via Python's exec() by a background cron job every 3-5 minutes, injected Python executes at the OS level. On default installations (PIALERT_WEB_PROTECTION = False) no credentials are required, matching the CVSS 9.8 network/no-privilege rating; there is no public exploit identified at time of analysis and the CVE is not in CISA KEV, but trivial complexity and full CIA impact make it a high-priority patch.
Stored cross-site scripting in the RELATE web courseware lets any enrolled student inject JavaScript that executes in an administrator's authenticated browser session, enabling full admin account takeover. The payload is planted via the freely editable first_name/last_name fields on the /profile/ page and fires when an admin opens the Participation list in the Django admin panel. No public exploit has been identified, but the root cause is confirmed in source and fixed upstream; with a CVSS of 8.7 and a scope-changing impact, this is a high-severity privilege-escalation issue.
Unauthorized file disclosure in Taipy 4.1.1 lets remote unauthenticated attackers read files outside an extension library's intended directory through the GUI ElementLibrary.get_resource() resource handler. The containment check used str.startswith() without a trailing separator, so a crafted request with traversal segments can resolve into a prefix-matching sibling directory on disk while still passing the flawed check. Impact is confined to confidentiality (file read), with no public exploit identified at time of analysis and no CISA KEV listing.
Authentication bypass in MaxKB (1Panel-dev) versions prior to 2.9.0 allows remote unauthenticated attackers to invoke webhook trigger endpoints and execute their bound tasks. The flaw stems from the WebhookAuth class unconditionally returning a successful authentication tuple, which Django REST Framework interprets as a valid identity, combined with no backend enforcement of per-trigger token requirements. No public exploit identified at time of analysis, but the trivial nature of the bypass and open-source visibility of the patch make exploitation straightforward for any attacker who can enumerate or guess trigger IDs.
Command injection in Vowpal Wabbit's GitHub Actions CI workflow allows an
Remote code execution in HuggingFace Transformers prior to 5.3.0 allows attackers to achieve arbitrary code execution on a victim's machine by publishing a malicious model whose config.json sets the `_attn_implementation_internal` field to an attacker-controlled Hub repository. When the victim calls the standard `AutoModelForCausalLM.from_pretrained()` API, the library silently downloads and executes Python kernels from that repository with the victim's privileges, bypassing the `trust_remote_code` safety gate. No public exploit is identified at time of analysis (EPSS 0.03%, SSVC exploitation: none), but the technical impact is total and the attack uses the documented, default usage pattern.
Arbitrary code execution in Docker Desktop's Model Runner on macOS allows any container on the Docker network to achieve RCE on the host by tricking the MLX inference backend into loading a Python file from an attacker-controlled OCI model registry. The MLX-LM library imports the file referenced by config.json's model_file field via importlib without any trust_remote_code gate, and the backend runs unsandboxed as the Docker Desktop user. Patched in Docker Desktop 4.71.0; no public exploit identified at time of analysis and EPSS is very low (0.01%), but the SSVC technical impact is rated total.
Arbitrary code execution in Docker Desktop's vllm-metal inference backend on macOS allows any container on the Docker network to trigger host-level RCE by pulling a malicious model from an OCI registry and requesting inference. The Docker Model Runner unconditionally sets trust_remote_code=True and runs without sandboxing, so AutoTokenizer.from_pretrained() loads attacker-controlled Python from the model and executes it as the Docker Desktop user. No public exploit identified at time of analysis; EPSS sits at 0.01% and SSVC marks exploitation as 'none' despite total technical impact.
Unauthenticated SQL injection in YesWiki's Bazar form-import path allows any remote visitor to inject arbitrary SQL into an INSERT statement and exfiltrate the entire database, including yeswiki_users.password hashes. Affects YesWiki 4.6.1, 4.6.2, and the doryphore-dev branch prior to 4.6.4. Publicly available exploit code exists (a working Python PoC is published in the GHSA advisory), though no public exploit identified in CISA KEV at time of analysis.
Unauthenticated cross-origin MCP tool invocation in Network-AI v5.4.4 allows a remote attacker to lure a victim to a malicious web page that silently invokes any of the 22 exposed MCP tools (including config_set, agent_spawn, blackboard_write, and token_create/revoke) against the victim's locally running MCP SSE server. The vulnerability stems from an empty default secret combined with a wildcard CORS policy, and publicly available exploit code exists in the GHSA advisory demonstrating end-to-end exploitation. No CISA KEV listing yet and EPSS data was not provided, but the published PoC and trivial attack mechanics make this a meaningful risk for any user running the default Docker deployment.
Arbitrary file write on the host in Boxlite sandbox service versions prior to 0.9.0 allows attackers to escape the OCI image extraction root via crafted symlink entries in layer tarballs, enabling remote code execution on the host (typically as root). Exploitation requires a user to pull and load a malicious OCI image distributed through registries such as DockerHub. Publicly available exploit code exists (vendor-published PoC); no public exploit identified in CISA KEV at time of analysis.
Sandbox escape in Boxlite versions prior to 0.9.0 lets untrusted code running inside the lightweight VM remount host-shared virtiofs directories from read-only to read-write, enabling arbitrary writes to host files that operators believed were protected. Because the container is granted all 41 Linux capabilities (including CAP_SYS_ADMIN), a trivial 'mount -o remount,rw' bypasses the client-side MS_RDONLY enforcement, and in AI-agent deployments this leads to host code execution by tampering with mounted code, virtualenvs, or credentials. Publicly available exploit code exists (working PoC published in the GHSA advisory) and the issue carries a CVSS 10.0 with scope change; no public exploit identified at time of analysis in CISA KEV.
Unsafe default code execution in InternLM LMDeploy (<=0.12.3) lets a malicious Hugging Face model repository run arbitrary Python on the host whenever a user loads it through any LMDeploy CLI (serve, calibrate, gptq, awq). The library hardcodes transformers.AutoConfig.from_pretrained(..., trust_remote_code=True) in get_model_arch and related helpers with no flag, env var, or warning to opt out, overriding HF Transformers' default-secure stance. No public exploit identified at time of analysis, and exploitation requires the user to load an untrusted repo, so risk is hardening-level rather than network-reachable RCE.
Arbitrary code execution in InternLM lmdeploy <= 0.12.3 occurs because trust_remote_code=True is hardcoded across HuggingFace model-loading call sites in lmdeploy/archs.py and lmdeploy/utils.py. An attacker who can influence the model_path passed to an lmdeploy serving process can point it at a malicious HuggingFace repository, causing Transformers to download and execute attacker-controlled Python code with the privileges of the serving daemon. Publicly available exploit code exists in the GHSA advisory, and an upstream fix has been merged via PR #4511 (fixed in 0.13.0).
Insecure deserialization in Apache Fory's PyFory (Python) library allows remote attackers to bypass DeserializationPolicy validation hooks via the ReduceSerializer, letting untrusted classes, functions, or module attributes be restored despite policy restrictions. All PyFory releases before 1.0.0 are affected when running Python-native mode with strict mode disabled, enabling attacker-controlled data to instantiate unsafe objects and achieve code execution. There is no public exploit identified at time of analysis and EPSS probability is very low (0.04%), but CVSS is 9.8 and SSVC rates technical impact as total with automatable exploitation.
Remote code execution in Hugging Face diffusers (Python package, versions < 0.38.0) is achievable via a TOCTOU race between two sequential Hub downloads inside DiffusionPipeline.from_pretrained, letting a malicious repo owner bypass the trust_remote_code guard and silently execute arbitrary Python during model loading. Exploitation requires user interaction (loading a malicious repo without pinning a revision) and high attack complexity due to a sub-second race window, but no public exploit beyond the reporter's PoC is identified at time of analysis. Affected users running diffusers <0.38.0 should upgrade to 0.38.0 where the issue is fixed.
Unauthenticated remote code execution in 9router (npm package) versions 0.4.30 through 0.4.36 allows network-adjacent attackers to execute arbitrary OS commands by chaining two unprotected API endpoints. The Next.js authentication middleware in src/proxy.js uses a narrow route allowlist that excludes /api/cli-tools/* and /api/mcp/*, letting an attacker register an arbitrary command via POST /api/cli-tools/cowork-settings and then trigger spawn() via GET /api/mcp/[plugin]/sse. Publicly available exploit code exists (PoC published with the GHSA advisory), with CVSS 10.0 reflecting maximum severity across confidentiality, integrity, and availability.
Arbitrary file write via path traversal in Mailpit's `dump --http` subcommand (versions < 1.30.0) allows any HTTP server impersonating a Mailpit instance to write attacker-controlled bytes to arbitrary paths outside the intended output directory. The attacker controls both the file path (via the message ID field in the JSON response) and the file contents (via the raw message body endpoint), enabling writes anywhere the dumping user has write permission - including cron jobs, shell startup files, and CI artifact directories. Publicly available exploit code exists (Python PoC published in GHSA-qx5x-85p8-vg4j); no confirmed active exploitation at time of analysis.
Server-side request forgery in the zrok Python SDK ProxyShare (pip/zrok 0.4.47 through 1.1.11) lets any viewer of a share coerce the proxy host into requesting arbitrary internal or loopback services and returns the response to the attacker. Because the Flask handler forwards the client-supplied path through urllib.parse.urljoin, an absolute URL in the request path overrides the share owner's fixed target host. No public exploit has been identified at time of analysis and EPSS is low (0.06%), but the CVSS 4.0 base score is 9.9 given unauthenticated network reach.
Information disclosure in Algernon web server versions 1.17.6 and earlier allows unauthenticated remote attackers to retrieve full server-side source code, including embedded secrets, by triggering runtime errors in Lua, Pongo2, Amber, or HTML template handlers. When Algernon is started with a single file path (e.g. `algernon page.po2`), single-file mode unconditionally forces debug mode on, activating the PrettyError renderer which returns absolute file paths and complete file contents in HTTP 200 responses. Crucially, the `--prod` hardening flag does not block this behavior for non-`.lua` extensions, and publicly available exploit code exists in the GHSA advisory.
Remote code execution in APScheduler (all versions through 3.10.x and 4.0.0a5) is achievable when applications deserialize attacker-controlled data via the bundled JSONSerializer or CBORSerializer. The unmarshal_object routine dynamically imports modules and invokes __setstate__ on arbitrary classes, letting an attacker pivot an untrusted payload into code execution; publicly available exploit code exists, though EPSS remains low at 0.06% (19th percentile).
Local privilege escalation to arbitrary code execution in MLflow versions prior to 3.11.0 stems from insecure temporary directory permissions (0o777 and 0o770) created by NFS and model-download helpers. Any local user sharing the filesystem - particularly on Databricks where NFS is enabled by default - can overwrite cloudpickle-serialized model artifacts and gain code execution when another user's process deserializes them via cloudpickle.load(). No public exploit is identified at time of analysis, and the issue is a continuation of CVE-2025-10279 which was only partially fixed.
Local file disclosure in NiceGUI versions <= 3.11.1 allows remote unauthenticated attackers to read arbitrary files accessible to the server process when applications pass user-controlled content to ui.restructured_text(). The flaw stems from Docutils being invoked without disabling file-insertion directives (include, csv-table :file:, raw :file:), enabling exfiltration of secrets, credentials, and source code. No public exploit identified at time of analysis, but the vendor advisory provides full directive-level proof patterns.
Remote denial-of-service in OpenTelemetry eBPF Instrumentation (OBI) versions 0.7.0 through 0.8.x allows unauthenticated attackers to crash the privileged instrumentation process by sending a crafted memcached storage command with an oversized `<bytes>` field. The integer overflow in the memcached text protocol parser produces a negative payload length that triggers a Go runtime panic in LargeBufferReader.Peek, halting telemetry collection until OBI is restarted. Publicly available exploit code exists in the GHSA-43g7-cwr8-q3jh advisory, but there is no public exploit identified beyond the PoC and the vulnerability is not listed in CISA KEV.
Remote code execution in the amazon-redshift-python-driver (versions prior to 2.1.14) allows a malicious or compromised Redshift server, or a man-in-the-middle attacker positioned on the network path, to execute arbitrary Python code on any client that connects. The root cause is unsafe use of Python's eval() against untrusted server-supplied data inside the vector_in() function. No public exploit identified at time of analysis, but the CVSS 4.0 base score of 9.3 and PR:N/UI:N vector make this a high-priority client-side supply-chain-style risk.
Denial of service in OpenTelemetry eBPF Instrumentation (OBI) versions prior to 0.9.0 allows remote attackers to crash the telemetry agent by sending a malformed Postgres BIND frame with an empty or unterminated portal name payload to any monitored service. The defect lives in OBI's passive Postgres protocol parser, where missing NUL-terminator validation causes a Go slice-bounds panic, halting telemetry collection on the affected node. Publicly available exploit code exists in the GHSA-pgvv-q3wf-mm9m advisory, though the issue is not listed in CISA KEV and EPSS data was not provided.
Remote code execution in ChromaDB Python (version 1.0.0 and later) allows unauthenticated attackers to execute arbitrary code on the server by submitting a malicious model repository with trust_remote_code enabled via the /api/v2/tenants/{tenant}/databases/{db}/collections endpoint. The flaw carries a maximum CVSS 4.0 score of 10.0 and was disclosed publicly by HiddenLayer; no public exploit identified at time of analysis, though detailed research has been published.
Unauthenticated remote code execution in SGLang (the LLM/multimodal generation serving runtime) affecting version 5.10 arises when the non-default `--enable-custom-logit-processor` flag is set, allowing attacker-supplied Python objects to be deserialized via `dill.loads()` and execute arbitrary code on the inference host. No CISA KEV listing exists and SSVC records exploitation as 'none', but a public technical write-up (antiproof.ai, 'Three RCEs in SGLang') details the flaw and SSVC marks it automatable with total technical impact. EPSS is modest at 0.32% (55th percentile), consistent with a serious-but-conditional (feature-gated) issue rather than mass exploitation.
Budibase's REST datasource integration before version 3.38.1 bypasses IP blacklist security controls through HTTP redirect following. Authenticated Builder-level users can exploit this to access cloud metadata services and internal databases by redirecting requests through attacker-controlled servers, potentially stealing AWS/GCP/Azure credentials. This vulnerability class was previously fixed in automation steps but the REST integration was overlooked, creating an inconsistent security posture.
Path traversal in Pipecat's development runner allows unauthenticated remote attackers to read arbitrary files when the --folder flag is used. The /files/{filename:path} endpoint fails to validate paths containing %2F-encoded directory separators, bypassing Starlette's URL normalization. Fixed in version 1.2.0 with no public exploit identified at time of analysis.
Authentication bypass in MLflow 3.9.0 and earlier allows unauthenticated remote attackers to access protected Job API and OpenTelemetry trace ingestion endpoints when the server runs with basic-auth enabled via uvicorn/ASGI. Attackers can submit jobs, read results, cancel operations, and inject trace data without credentials. The FastAPI permission middleware incorrectly enforced authentication only on /gateway/ routes, leaving /ajax-api/3.0/jobs/* and /v1/traces unprotected due to architectural mismatch between Flask and FastAPI authentication mechanisms. Fixed in version 3.10.0 with GitHub commit bb62e77 adding proper validators for all FastAPI routes.
python-utcp CLI subprocess environment passes all process-level secrets to every tool call. When chained with CVE-2026-45369 command injection, remote authenticated attackers with low-privilege LLM tool access can exfiltrate AWS credentials, API keys, database URLs, and other environment variables in a single HTTP request. Patch available in version 1.1.2 (NVD references 1.1.3 as fixed version). GitHub security advisory confirms proof-of-concept demonstrating credential theft via env dump to attacker-controlled endpoint.
Command injection in python-utcp allows remote attackers to execute arbitrary shell commands on Unix and Windows systems when user-controlled tool arguments are processed by the CLI communication protocol module. The _substitute_utcp_args method in cli_communication_protocol.py directly embeds unsanitized user input into bash or PowerShell commands without escaping, enabling full remote code execution. Vendor-released patch available in version 1.1.2 with shell-quoting mitigation (shlex.quote on Unix, single-quoted literals on Windows). CVSS 8.3 indicates high complexity and required user interaction, but scope change enables container/sandbox escape scenarios. No public exploit code or CISA KEV listing identified at time of analysis, though detailed proof-of-concept exists in the GitHub security advisory demonstrating data exfiltration via curl.
Multiple concurrent LDAP or OAuth first-login requests on a freshly deployed Open WebUI instance can all receive administrator privileges through a TOCTOU race condition in role assignment logic. The vulnerability affects deployments using LDAP or OAuth authentication on instances with no existing users. While the regular signup handler was explicitly patched for this race condition in earlier code ('Insert with default role first to avoid TOCTOU race'), the LDAP and OAuth authentication paths were never updated with the same fix. Vendor-released patch available in version 0.9.0 (April 2026). No active exploitation confirmed (not in CISA KEV), though publicly available exploit code exists per GitHub advisory GHSA-h3ww-q6xx-w7x3. CVSS 8.1 (High) reflects network attack vector but requires high attack complexity (precise timing of concurrent requests during narrow first-deployment window).
Open WebUI versions through 0.8.11 allow authenticated users to execute arbitrary Python code in the Jupyter container by bypassing the ENABLE_CODE_EXECUTION=false configuration flag. The /api/v1/utils/code/execute endpoint fails to enforce the admin-configured feature gate (CWE-863: Incorrect Authorization), enabling any verified user to run code even when administrators believe execution is disabled. The vulnerability is confirmed by vendor POC (verified 2026-03-25) demonstrating successful code execution, file access, and SSRF to internal Docker services despite explicit admin configuration disabling the feature. Vendor-released patch available in v0.8.12 (commit 6d736d3c5) enforces the configuration check before dispatching code to Jupyter.
Broken object-level authorization in Open WebUI versions ≤0.8.12 allows any authenticated user to permanently delete files owned by other users when those files are referenced in any shared chat. The has_access_to_file() authorization function unconditionally grants access through its shared-chat branch, failing to validate both the requesting user's identity and the operation type (read vs. write). File UUIDs, which would otherwise prevent enumeration attacks, are exposed via the knowledge base API endpoint GET /api/v1/knowledge/{id}/files to any user with read access. This affects all default Docker deployments where chat sharing is enabled. Vendor-released patch available in v0.9.0 (commit 2e52ad8ff). No active exploitation confirmed (not in CISA KEV). CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:U/C:H/I:H/A:H scores 8.0, though real-world impact extends beyond confidentiality to permanent data destruction with no recovery mechanism.
URL parser mismatch in Open WebUI allows authenticated users to bypass SSRF protections and access internal network resources. The validate_url function uses Python's urlparse library to extract hostnames for validation, while the requests library handles actual HTTP requests. These libraries disagree on parsing URLs containing backslash characters (e.g., http://127.0.0.1:6666\@1.1.1.1), allowing attackers to craft URLs that pass validation as external addresses but resolve to internal hosts. Exploitation requires low-privilege authentication but no user interaction, enabling access to cloud metadata endpoints and internal services. Fixed in version 0.9.5 per GitHub advisory GHSA-8w7q-q5jp-jvgx.
Broken object-level authorization in Open WebUI allows any authenticated low-privilege user to enumerate and terminate background tasks belonging to other users via GET /api/tasks and POST /api/tasks/stop/{task_id}. Attackers can disrupt system-wide chat generation and background processing by continuously canceling active tasks across the multi-user instance. Publicly available exploit code exists. Vendor-released patch in v0.9.0 restricts global task endpoints to admin-only and introduces a scoped /api/tasks/chat/{chat_id}/stop endpoint for legitimate user-owned task termination. CVSS 7.1 (AV:N/AC:L/PR:L/UI:N) reflects network-accessible, low-complexity exploitation requiring only authenticated low-privilege access, with high availability impact and low confidentiality impact from task enumeration.
Insecure Direct Object Reference (IDOR) in Open WebUI's retrieval API allows authenticated users to bypass knowledge base access controls and directly access, modify, or delete other users' private knowledge bases by supplying the target UUID as a collection name. The authorization gap affects seven endpoints: two read endpoints (/query/doc, /query/collection) permit exfiltration of private knowledge base content, while five write endpoints (/process/text, /process/file, /process/files/batch, /process/web, /process/youtube) enable content injection, poisoning, or complete data destruction via overwrite. Affects Open WebUI <= 0.9.4; fixed in v0.9.5 via PR #22109. EPSS data not available; no confirmed active exploitation (CVSS 7.5 reflects AC:H due to UUID prerequisite, but UUIDs leak through multiple channels per researcher analysis).
Privilege escalation in Open WebUI allows authenticated users with 'write' access grants on tools to execute arbitrary server-side Python code without the required workspace.tools permission. The tool update endpoint (POST /api/v1/tools/id/{id}/update) fails to enforce the workspace.tools authorization check that gates code execution, allowing users explicitly denied code execution capabilities to bypass this security boundary. This breaks Open WebUI's documented trust model where workspace.tools permission is intentionally disabled by default and 'equivalent to giving them shell access to the server.' Exploitation achieves root code execution (PID 1) in default Docker deployments, enabling extraction of secrets (WEBUI_SECRET_KEY, API keys), database access, and filesystem read/write. Confirmed by GitHub security advisory GHSA-p4fx-23fq-jfg6. No public exploit or KEV listing at time of analysis, but detailed proof-of-concept with Burp Collaborator confirmation exists in the advisory.
Server-Side Request Forgery (SSRF) in Open WebUI versions ≤0.8.12 allows authenticated users with OAuth access to force the server to make HTTP requests to arbitrary internal resources and exfiltrate complete response data. Exploitation requires OAuth-enabled deployments with ENABLE_OAUTH_SIGNUP=true or OAUTH_UPDATE_PICTURE_ON_LOGIN=true. An attacker controls the OAuth provider's 'picture' claim URL, triggering server-side HTTP requests to cloud metadata services (AWS IMDS), localhost services (Redis, Elasticsearch), or internal network endpoints. The full response is base64-encoded and stored in the user's profile_image_url field, enabling complete data exfiltration. Fixed in version 0.9.0 per GitHub advisory GHSA-24c9-2m8q-qhmh. EPSS data not available; no CISA KEV listing indicates limited widespread exploitation, though publicly available proof-of-concept exists in the GitHub advisory.
Server-Side Request Forgery in Open WebUI's `validate_url()` function allows authenticated attackers to reach internal IPv4/IPv6 addresses, bypassing security controls via three distinct flaws: the validators library silently fails on IPv6 private-address checks (raising ValidationError which evaluates as falsy), IPv4-mapped IPv6 addresses (::ffff:10.0.0.1) evade IPv4 filtering entirely, and multiple IANA-reserved IPv4 ranges (0.0.0.0/8, 100.64.0.0/10, 192.0.0.0/24, 198.18.0.0/15, 203.0.113.0/24) remain unblocked. The vulnerability persists in the RAG web search, image editing, and other endpoints despite an earlier incomplete remediation attempt (CVE-2025-65958), enabling exfiltration of AWS IMDSv1 credentials and access to localhost-bound services. Publicly available exploit code exists (demonstrated POC in advisory), affecting Open WebUI ≤0.8.12 with fix released in version 0.9.0.
Stored cross-site scripting (XSS) in Open WebUI ≤0.9.2 allows authenticated users with default speech-to-text permissions to upload polyglot WAV+HTML files through the audio transcription endpoint, achieving code execution in victim browsers and enabling full account takeover including administrator sessions. The vulnerability chains insecure file extension handling with unrestricted Content-Type serving and non-HttpOnly JWT storage to weaponize a single-click attack. Publicly available exploit code exists with video demonstration; no active exploitation confirmed by CISA KEV at time of analysis. CVSS 8.7 (High) reflects changed scope (S:C) and user interaction requirement, but real-world risk is elevated because the vulnerable permission defaults to enabled and the attack yields immediate admin-level access in typical deployments.
Cleartext HMAC signing key exposure in Amazon SageMaker Python SDK versions <2.257.2 and <3.8.0 enables authenticated attackers with SageMaker describe API and S3 write permissions to forge model artifact integrity signatures and achieve remote code execution in inference containers. AWS released patches in v2.257.2 and v3.8.0 with security fixes addressing Triton HMAC key exposure and missing integrity checks. EPSS data not available; no CISA KEV listing or public POC identified at time of analysis, suggesting limited exploitation activity despite high CVSS score.
Authenticated server-side request forgery in ApostropheCMS allows low-privilege users to force the server to fetch arbitrary internal URLs through the rich-text widget import flow. Attackers with content editing permissions can exfiltrate internal data by crafting malicious image tags that trigger server-side fetch operations, with image-compatible responses being persisted and re-hosted by the application. Publicly available exploit code exists (full Python PoC published in GitHub advisory GHSA-pr28-mf3q-qpg6), enabling immediate weaponization. All versions through 4.29.0 are affected with no vendor-released patch identified at time of analysis, creating sustained exposure for organizations running this popular Node.js CMS.
Remote code execution in HuggingFace Diffusers library (versions < 0.38.0) allows attackers to execute arbitrary Python code when victims load malicious pipelines from Hugging Face Hub repositories. The vulnerability bypasses the trust_remote_code=True safeguard through a type coercion flaw where None values are interpolated as 'None.py' filenames. Attackers can achieve silent code execution by publishing repositories containing a malicious None.py file alongside legitimate-looking configuration, requiring only that victims call DiffusionPipeline.from_pretrained() on the attacker's repository. EPSS data not available; no public exploit identified at time of analysis. Vendor-released patch: version 0.38.0.
Privilege escalation in wger fitness manager allows gym trainers to impersonate gym managers via session-chain attack. An authenticated trainer exploits flawed session-flag logic in the trainer-login endpoint to bypass permission checks - first switching into a low-privilege user, then leveraging the inherited 'trainer.identity' session flag to hop into manager accounts. Publicly available proof-of-concept demonstrates complete takeover of gym administration with CVSS 8.1 (network-accessible, low complexity). No vendor patch confirmed at time of analysis; vulnerability actively disclosed by wger-project GitHub advisory GHSA-9qpr-vc49-hqg2. EPSS score not available, not in CISA KEV. Root cause is CWE-269 (improper privilege management) in core/views/user.py lines 169-178.
Insecure Direct Object Reference in wger fitness platform exposes any authenticated user's complete workout history, session notes, and training statistics through template routine API endpoints. Attackers with free accounts enumerate public template routine IDs and retrieve owners' private health data including workout notes, weights, repetitions, and performance metrics via /api/v2/routine/{id}/logs/ and /api/v2/routine/{id}/stats/ endpoints. Detailed proof-of-concept with Python exploit confirms trivial exploitation against wger <= 2.5.0a2. CVSS 7.5 rates this High severity, but NOTE: vector PR:N appears inconsistent with authenticated-only access described - attackers need valid credentials, suggesting actual vector should be PR:L. EPSS data not available. No CVE KEV listing or public exploit repositories identified beyond GitHub advisory disclosure. Patch status unconfirmed - GitHub advisory references fix commit but no released version number provided in available data.
Local code execution in the claude-code-cache-fix npm package (v3.5.0 and v3.5.1) lets attacker-controlled filesystem path names run arbitrary Python inside a victim's Claude Code process. The bundled tools/quota-statusline.sh interpolates Claude Code's statusline hook stdin — which reflects user-controlled paths such as cwd, workspace.current_dir, workspace.project_dir, and transcript_path — directly into a Python triple-quoted literal, so a directory name containing the byte sequence ''' closes the literal early and executes following bytes as Python at the user's privilege on every statusline redraw. A working injection payload is publicly available exploit code (published in the GHSA advisory and the T6/T7 regression tests); the issue is not listed in CISA KEV and no EPSS score was provided.
Unsafe deserialization in LangSmith SDK's prompt pull methods allows remote attackers to execute server-side request forgery (SSRF) and redirect LLM traffic to attacker-controlled infrastructure when applications pull public prompts from LangSmith Hub. The SDK deserializes untrusted prompt manifests containing serialized LangChain objects with attacker-controlled constructor arguments, including malicious base_url configurations, custom headers, and secret references. Exploitation requires user interaction (developers must call pull_prompt with a malicious owner/name identifier), but no authentication is required to publish malicious prompts to the public Hub. Vendor-released patches in Python >= 0.8.0 and JS/TS >= 0.6.0 now block public prompt pulling by default, requiring explicit opt-in via dangerously_pull_public_prompt flag. EPSS data not available; no CISA KEV listing or public exploit identified at time of analysis.
Denial-of-service via memory exhaustion affects UltraJSON (ujson) versions 5.12.0 and earlier, where the ujson.dump() function fails to release (Py_DECREF) the serialized JSON string object when the target file-like object's write() method raises an exception. Each failed write leaks the full serialized payload, so an attacker who can repeatedly trigger write failures - for example by disconnecting mid-response from a web server that streams JSON via ujson.dump() - can drive linear, unbounded memory growth. A proof-of-concept exists in the vendor advisory (GHSA-c38f-wx89-p2xg), but there is no public exploit identified as weaponized and no active exploitation; EPSS is very low at 0.04% (12th percentile).
Sandbox escape in Heym's custom Python tool executor (versions before 0.0.21) allows authenticated workflow authors to break out of the restricted execution environment using Python object-graph introspection, recover the unrestricted __import__ function, and load blocked modules such as os and subprocess. Once free of the sandbox, an attacker executes arbitrary host commands as the backend service user and reads inherited environment variables holding database credentials and encryption keys. Publicly available exploit code exists (SSVC 'poc'), though the EPSS probability is very low at 0.04% and it is not listed in CISA KEV.
Sandbox escape in OpenClaude (npm package openclaude) versions before 0.5.1 allows a prompt-injected LLM to disable host sandboxing by setting the model-controlled `dangerouslyDisableSandbox: true` flag in any Bash tool_use call, yielding full unsandboxed command execution on the host. CVSS 4.0 scores this 9.3 Critical (AV:N/AC:L/PR:N/UI:N, VC/VI/VA:H); no public exploit identified at time of analysis beyond the reporter's PoC, but the upstream fix has been merged. The flaw is especially severe because it is reachable under default settings (`allowUnsandboxedCommands` defaults to true).
Arbitrary code execution via torch-checkpoint-shrink.py script in ml-engineering project allows remote attackers to execute malicious Python code by providing crafted PyTorch checkpoint files. The vulnerability stems from insecure deserialization where torch.load() processes .pt files without the weights_only=True safeguard, enabling pickle-based arbitrary object instantiation. Despite a critical CVSS 9.8 score, EPSS probability is low (0.06%, 19th percentile) and no public exploit or active exploitation is confirmed, suggesting limited real-world targeting to date. SSVC assessment indicates total technical impact with automatable exploitation potential, making this a priority for organizations using ml-engineering scripts in production environments.
Remote code execution in Ludwig framework ≤0.10.4 allows unauthenticated network attackers to execute arbitrary code by supplying a malicious PyTorch model file to the ludwig serve endpoint. The vulnerability stems from unsafe deserialization in the model loading component, which uses torch.load() without the weights_only=True safety parameter. With CVSS 9.8 (critical network vector, no authentication required) but only 0.02% EPSS, this represents a high-severity issue in vulnerable deployments, though widespread exploitation has not been observed. No CISA KEV listing or public POC identified at time of analysis.
Arbitrary code execution in Ludwig framework ≤0.10.4 occurs when attackers supply malicious pickle files to the predict() method, which deserializes untrusted data without validation using pandas.read_pickle(). Remote unauthenticated attackers can achieve full system compromise by exploiting the automatic file format detection mechanism that processes .pkl files through Python's unsafe pickle module. EPSS score of 0.06% (19th percentile) suggests low current exploitation likelihood despite the critical CVSS 9.8 rating, though no public exploit code or active exploitation has been identified at time of analysis.
Remote code execution in Mamba language model framework (through version 2.2.6) allows unauthenticated attackers to execute arbitrary Python code by publishing malicious models on HuggingFace Hub. When victims call MambaLMHeadModel.from_pretrained() on a weaponized model repository, insecure pickle deserialization executes attacker-controlled code in the context of the victim's process. Despite the critical CVSS 9.8 score and network attack vector requiring no authentication, EPSS probability remains extremely low (0.02%, 5th percentile), suggesting limited real-world exploitation to date. No CISA KEV listing or public POC identified at time of analysis.
Remote code execution in Optimate's neural_magic_training.py script allows authenticated attackers to execute arbitrary code via malicious PyTorch model files. The vulnerability stems from unsafe deserialization when loading model state dictionaries without PyTorch's weights_only=True security flag, enabling pickle-based arbitrary object execution. With an EPSS score of 0.06% and no confirmed exploitation, this represents a moderate risk primarily in environments where users can upload or specify model files.
Arbitrary code execution in optimate's neural_magic_training.py allows remote attackers to execute Python code by supplying a malicious directory path containing a crafted module.py file. The _load_model() function directly executes file contents via Python's exec() without validation. CVSS 9.8 reflects network vector with no authentication, but EPSS score of 0.02% (5th percentile) indicates very low observed exploitation probability. No active exploitation confirmed (not in CISA KEV). Vulnerability exists in commit a6d302f912b481c94370811af6b11402f51d377f from July 2024. Affects organizations using optimate for neural network model optimization.
Remote code execution in PySyft Datasite/Server versions 0.9.5 and earlier allows unauthenticated attackers to execute arbitrary Python code on the server through the function submission mechanism. The vulnerability stems from insufficient validation and sandboxing of user-submitted Python functions decorated with @sy.syft_function(), which are executed using unsafe exec() and eval() calls after approval. With an EPSS score of 0.04% and no current KEV listing, this appears to be a high-severity vulnerability without confirmed active exploitation.
Remote code execution in Snorkel machine learning library (≤v0.10.0) occurs when users load untrusted model files via MultitaskClassifier.load(). The vulnerability exploits insecure Python object deserialization through torch.load(), allowing attackers to embed malicious code in model weight files that executes upon loading. EPSS score of 0.06% (19th percentile) suggests low observed exploitation probability in the wild, though SSVC framework indicates total technical impact once exploited. No public exploit code or active exploitation confirmed at time of analysis, but exploitation requires only that a data scientist or ML engineer load a malicious .pkl model file.
Insecure deserialization in Optimate's neural_magic_training.py script enables remote code execution when loading PyTorch model files. The _load_model() function uses torch.load() without the weights_only=True security parameter, allowing attackers with low privileges to execute arbitrary Python code by providing malicious .pt or .pth files via the --model command-line argument. EPSS indicates low exploitation probability at 0.06% with no active exploitation confirmed.