AI / ML

NVIDIA NeMo Framework contains a vulnerability where an attacker could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, denial of service, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework contains a vulnerability where an attacker could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, denial of service, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework contains a vulnerability where an attacker could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, denial of service, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework for all platforms contains a vulnerability in a voice-preprocessing script, where malicious input created by an attacker could cause a code injection. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework for all platforms contains a vulnerability in the ASR Evaluator utility, where a user could cause a command injection by supplying crafted input to a configuration parameter. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework contains a vulnerability where malicious data could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 8.0 HIGH]

NVIDIA NeMo Framework contains a vulnerability where an attacker could cause remote code execution in distributed environments. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework contains a vulnerability where an attacker could cause remote code execution by loading a maliciously crafted file. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA Megatron Bridge contains a vulnerability in a data shuffling tutorial, where malicious input could cause a code injection. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA Megatron Bridge contains a vulnerability in a data merging tutorial, where malicious input could cause a code injection. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 7.8 HIGH]

NVIDIA NeMo Framework contains a vulnerability where malicious data created by an attacker could cause code injection. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering. [CVSS 7.8 HIGH]

Insufficient input validation in the Gutenberg Blocks with AI by Kadence WP plugin allows authenticated contributors and above to perform server-side request forgery against GetResponse API endpoints, potentially exposing sensitive data like contacts and campaigns stored on the site. The vulnerability stems from overly permissive access controls that grant the dangerous `endpoint` parameter manipulation to users with only Contributor-level privileges instead of requiring administrator access. Attackers can also extract the site's stored GetResponse API credentials from request headers during exploitation.

JeecgBoot 3.9.1's RAG knowledge controller fails to properly validate ZIP file imports, allowing authenticated remote attackers to trigger unsafe deserialization with public exploit code available. The vulnerability requires authentication and complex attack execution but could enable information disclosure or integrity compromise. No patch is currently available from the vendor.

Server-side request forgery in Deepaudit versions up to 3.0.3 allows authenticated remote attackers to manipulate the IP Address Handler component in the embedding configuration endpoint, potentially enabling them to perform arbitrary network requests from the affected server. The vulnerability requires valid credentials but no user interaction, affecting the AI/ML product's backend services. Upgrading to version 3.0.4 or later resolves this issue.

MindsDB versions up to 25.14.1 contain a server-side request forgery vulnerability in the file upload functionality that allows authenticated remote attackers to forge requests to internal or external systems. Public exploit code exists for this vulnerability, and affected organizations should apply patch 74d6f0fd4b630218519a700fbee1c05c7fd4b1ed or upgrade to a patched version immediately.

Unauthenticated API access in Milvus vector database before 2.5.27/2.6.10. TCP port 9091 exposed by default without authentication. EPSS 0.32% with PoC and patch available.

Cursor versions before 2.5 allow sandbox escape through improper .git configuration file protections, enabling malicious prompts or agents to write git hooks that execute arbitrary code when git commands are triggered. An attacker can achieve remote code execution without user interaction since git automatically executes these hooks, potentially compromising systems where Cursor is used for AI-assisted development. A patch is available in version 2.5.

FastGPT's web and HTTP data acquisition nodes fail to properly validate internal network addresses, allowing unauthenticated remote attackers to bypass network isolation controls and access sensitive internal resources. This vulnerability affects FastGPT versions prior to 4.14.7 and requires user interaction to exploit. The vulnerability has a CVSS score of 5.4 and currently has no available patch.

Remote code execution in AutoGPT prior to version 0.6.48 allows authenticated users to execute arbitrary Python code on the backend server by embedding a disabled BlockInstallationBlock within a workflow graph, bypassing validation controls that only checked the disabled flag at direct execution endpoints. An attacker with valid credentials can exploit this to gain full control over the backend system and automate malicious workflows. The vulnerability has been patched in version 0.6.48 and all users should upgrade immediately.

Crawl4AI versions before 0.8.0 allow unauthenticated remote attackers to read arbitrary files from the server through file:// URL handling in Docker API endpoints (/execute_js, /screenshot, /pdf, /html), enabling exposure of sensitive configuration files, credentials, and environment variables. The vulnerability affects Docker deployments and AI/ML applications using the affected library, with no patch currently available.

Remote code execution in Crawl4AI Docker API before 0.8.0 via hooks parameter. The /crawl endpoint accepts Python code in hooks that executes on the server. EPSS 0.28%.

Keras versions 3.0.0 through 3.13.1 are vulnerable to arbitrary file read through malicious .keras model files that abuse HDF5 external dataset references, enabling unauthenticated remote attackers to disclose sensitive local files. This high-severity vulnerability affects all supported platforms and currently has no available patch. An attacker can exploit this by crafting a specially formatted model file that, when loaded by a Keras application, reads arbitrary files from the system.

Dify versions prior to 1.13.0 contain a stored cross-site scripting vulnerability in the chat frontend's echarts integration that executes malicious JavaScript payloads embedded in user or LLM-generated inputs. An attacker can exploit this to perform actions in the context of other users' browsers, potentially stealing session tokens or conducting phishing attacks. Public exploit code exists for this vulnerability, though a patch is available in version 1.13.0 and later.

RecursiveUrlLoader in LangChain Community prior to 1.1.14 uses weak string-based URL validation that allows attackers to bypass the preventOutside crawling restriction by crafting domains with matching prefixes, potentially exposing the crawler to malicious or internal infrastructure endpoints. An attacker controlling a crawled webpage could inject links to cloud metadata services or private IP ranges, which the crawler would follow without validation, leading to information disclosure.

Unsafe deserialization in DiskCache Python library through 5.6.3. Uses pickle by default, allowing attackers with cache directory write access to execute arbitrary code.

LangChain is a framework for building agents and LLM-powered applications. Prior to 1.2.11, the ChatOpenAI.get_num_tokens_from_messages() method fetches arbitrary image_url values without validation when computing token counts for vision-enabled models. This allows attackers to trigger Server-Side Request Forgery (SSRF) attacks by providing malicious image URLs in user input. This vulnerability...

AutoGPT platform versions before 0.6.32 contain a regular expression denial of service vulnerability in the Code Extraction Block due to overlapping quantifiers that cause catastrophic backtracking when processing whitespace-heavy inputs. Authenticated attackers can exploit this by submitting malicious input with long sequences of spaces to trigger excessive regex processing, causing the service to become unavailable. Public exploit code exists for this vulnerability, and a patch is available in version 0.6.32 and later.

Unauthenticated access to the FastGPT plugin API endpoint (FastGPT/api/plugin/xxx) in versions 4.14.0 through 4.14.5 allows remote attackers to disrupt plugin functionality and cause loss of plugin installation state without authentication. The vulnerability affects the AI/ML platform's plugin system availability and integrity, though sensitive data such as cryptographic keys are not exposed. A patch is available in version 4.14.5-fix.

Authenticated users can exploit a race condition in GitHub Copilot and Visual Studio Code to execute arbitrary code remotely by manipulating file state between verification and use. This vulnerability affects users with network access to these development tools and requires user interaction to trigger. No patch is currently available to address this high-severity flaw.

GitHub Copilot and Visual Studio Code are vulnerable to command injection attacks that allow unauthenticated attackers to bypass security features over the network through improper neutralization of special command elements. The vulnerability requires user interaction to exploit and could enable attackers to execute arbitrary commands with high impact on confidentiality, integrity, and availability. No patch is currently available for this issue.

GitHub Copilot is vulnerable to command injection attacks that enable remote code execution without requiring authentication or user interaction beyond a click. An attacker can exploit this network-accessible vulnerability to execute arbitrary commands on affected systems. No patch is currently available for this high-severity vulnerability.

GitHub Copilot and Visual Studio 2022 contain a command injection vulnerability that allows authenticated users to execute arbitrary commands through improper sanitization of special elements. An attacker with valid credentials can leverage user interaction to escalate privileges and gain elevated access across the network. No patch is currently available for this vulnerability.

Remote code execution in GitHub Copilot and Visual Studio 2022 via command injection allows unauthenticated attackers to execute arbitrary code over the network with user interaction. The vulnerability stems from improper sanitization of special elements in commands, enabling attackers to break out of intended command contexts and inject malicious payloads. No patch is currently available for this high-severity issue affecting both development environments.

Uncontrolled search path for some AI Playground before version 2.6.1 beta within Ring 3: User Applications may allow an escalation of privilege. Unprivileged software adversary with an authenticated user combined with a high complexity attack may enable escalation of privilege. This result may potentially occur via local access when attack requirements are present without special internal knowl...

Authenticated attackers can inject malicious JavaScript into Flowring's AgentFlow platform that persists and executes in other users' browsers when they load affected pages, potentially compromising user sessions and data. This stored cross-site scripting vulnerability affects the AI/ML and Agentflow products and requires user interaction to trigger, though no patch is currently available.

Reflected XSS in AgentFlow enables unauthenticated attackers to inject malicious JavaScript that executes in victims' browsers during phishing campaigns, potentially compromising user sessions and data. The vulnerability affects the AI/ML platform with no patch currently available, requiring users to rely on defensive measures such as email filtering and user awareness training.

Agentflow versions up to - is affected by unrestricted upload of file with dangerous type (CVSS 8.8).

Missing authentication in Flowring Agentflow allows unauthenticated attackers to read, modify, and delete data. Second auth bypass CVE.

Stored cross-site scripting in the Fluent Forms WordPress plugin's AI Form Builder module (versions up to 6.1.14) enables authenticated subscribers to inject malicious scripts that execute for all users viewing affected forms through missing authorization checks, leaked nonces, and insufficient input sanitization of AI-generated content. An attacker with subscriber-level access can exploit this to perform actions on behalf of administrators or steal sensitive information from form viewers. The vulnerability affects WordPress installations using this plugin and has no patch currently available.

LangSmith Client SDKs for Python and AI/ML platforms are susceptible to server-side request forgery through malicious HTTP baggage headers that allow attackers to redirect trace data exfiltration to attacker-controlled endpoints. An unauthenticated attacker can inject arbitrary api_url values during distributed tracing operations, causing the SDK to send sensitive trace data outside the intended infrastructure. No patch is currently available for this medium-severity vulnerability.

Python code execution through Pyodide in the mcp-run-python library lacks isolation from the JavaScript environment, enabling attackers to manipulate the JS runtime and hijack MCP server functionality. This allows adversaries to perform malicious operations including tool shadowing and potential server compromise through crafted Python payloads. No patch is available as the project is archived.

Pydantic-AI's MCP Run Python tool uses an insufficiently restrictive Deno sandbox configuration that permits Python code to access the host's localhost interface, enabling Server-Side Request Forgery (SSRF) attacks. An attacker can exploit this to probe or interact with services running on the local machine that should be isolated from external access. The archived project status means no patch is expected to be released.

GitLab AI Gateway's Duo Workflow Service has a CVSS 9.9 server-side template injection enabling code execution through the AI workflow system.

Command injection in Xcode MCP Server's LLDB integration allows authenticated network attackers to execute arbitrary commands by manipulating the args parameter in the registerXcodeTools function. Public exploit code exists for this vulnerability, increasing the practical risk to organizations using affected versions. Users should apply the available patch to remediate this medium-severity flaw affecting the AI/ML tooling component.

OS command injection in XixianLiang HarmonyOS-mcp-server 0.1.0 allows authenticated remote attackers to execute arbitrary commands through unsanitized input to the input_text function. Public exploit code exists for this vulnerability, and no patch is currently available. An attacker with valid credentials can exploit this over the network to achieve remote code execution with limited complexity.

Maigret Mcp Server versions up to 1.0.12. contains a vulnerability that allows attackers to command injection (CVSS 6.3).

Path traversal in JeecgBoot's Retrieval-Augmented Generation Module (versions up to 3.9.0) allows authenticated remote attackers to access arbitrary files through manipulation of the filePath parameter in the /airag/knowledge/doc/edit endpoint. Public exploit code exists for this vulnerability, and no patch is currently available from the vendor despite early notification.

Enclave versions up to 2.10.1 is affected by loop with unreachable exit condition (infinite loop) (CVSS 8.8).

DeepAudit is a multi-agent system for code vulnerability discovery. [CVSS 6.5 MEDIUM]

Arbitrary file append vulnerability in Qdrant vector database versions 1.9.3 through 1.15.x allows authenticated users with minimal read-only privileges to write to arbitrary files through an unsanitized log file path parameter in the /logger endpoint. Public exploit code exists for this vulnerability, enabling attackers to corrupt system files or inject malicious content with high impact to confidentiality, integrity, and availability. The issue is resolved in version 1.16.0.

OpenClaw prior to version 2026.1.20 allows local unauthenticated attackers to execute arbitrary commands as the gateway user by exploiting the WebSocket API to inject malicious command paths through the config.apply function. The vulnerability stems from insufficient validation of the cliPath parameter, which is subsequently used for command discovery without proper sanitization. No patch is currently available for affected versions.

Microsoft Semantic Kernel SDK has a CVSS 9.9 path traversal vulnerability enabling AI agents to access arbitrary files outside their intended scope.

Pydantic AI versions 0.0.26 through 1.55.x contain a server-side request forgery vulnerability in URL download functionality that allows remote attackers to make arbitrary HTTP requests to internal network resources when applications process untrusted message history. Public exploit code exists for this vulnerability, which could enable attackers to access internal services or cloud credentials. Applications must upgrade to version 1.56.0 or later to remediate the issue.

Pydantic AI versions 1.34.0 through 1.50.x contain a path traversal vulnerability in the web UI that allows unauthenticated attackers to inject arbitrary JavaScript by manipulating the CDN version parameter in a malicious URL. When a victim visits the crafted link, attacker-controlled code executes in their browser, enabling theft of chat history and other sensitive client-side data. No patch is currently available.

Unauthenticated attackers can exploit arbitrary attribute access in MCP Salesforce Connector versions prior to 0.1.10 to extract sensitive Salesforce authentication tokens. This vulnerability requires only network access with no user interaction, enabling complete disclosure of credentials used for Salesforce API integration. Organizations using affected versions should upgrade to 0.1.10 immediately.

Claude Code prior to version 2.1.2 has a CVSS 10.0 sandbox escape in the bubblewrap sandboxing mechanism, allowing code execution outside the intended sandbox boundary.

Claude Code versions prior to 2.0.55 insufficiently validate piped sed commands, permitting authenticated users to circumvent file write protections and deposit files in restricted directories including .claude folders and locations outside project scope. An attacker with access to the "accept edits" feature can exploit this to write malicious content to sensitive areas of the system. A patch is available in version 2.0.55 and later.

Claude Code prior to version 2.0.57 failed to properly validate MCP tool inputs, allowing malicious MCP servers to inject commands through tool responses.

OrcaStatLLM Researcher is an LLM Based Research Paper Generator. [CVSS 5.4 MEDIUM]

Insufficient authorization checks in Ansible Lightspeed API conversation endpoints allow authenticated users to access and modify conversations belonging to other users. An attacker with valid credentials can exploit this to read sensitive conversation data and manipulate AI-generated outputs from other users' sessions. No patch is currently available.

AutoGPT is a platform that allows users to create, deploy, and manage continuous artificial intelligence agents that automate complex workflows. Prior to autogpt-platform-beta-v0.6.32, there is a DoS vulnerability in ReadRSSFeedBlock. [CVSS 6.5 MEDIUM]

AutoGPT platform versions prior to v0.6.46 expose API keys and authentication secrets in application logs due to insecure logging of decrypted credentials across three Stagehand integration blocks. Authenticated users can access these plaintext secrets through log files, enabling credential theft and unauthorized access to integrated services. Public exploit code exists for this vulnerability, though a patch is available in v0.6.46 and later.

AutoGPT has a second SSRF vulnerability (CVSS 9.8) in a different endpoint, providing an additional path to access internal network resources.

AutoGPT has a Server-Side Request Forgery vulnerability (CVSS 9.8) allowing unauthenticated attackers to make the AI platform access internal network resources.

Langroid LLM framework prior to 0.5 has a code injection vulnerability (CVSS 9.6) allowing attackers to execute arbitrary code through the AI agent system.

OpenClaw versions prior to 2026.1.30 suffer from a path traversal vulnerability in the isValidMedia() function that permits authenticated agents to read arbitrary files on the system by crafting malicious MEDIA output directives. An attacker with agent access can leverage this flaw to exfiltrate sensitive data accessible to the application process. Public exploit code exists for this vulnerability, and no patch is currently available.

OpenClaw AI assistant versions prior to 2026.1.29 contain two command injection vulnerabilities: unescaped user input in SSH project paths allows remote code execution on SSH hosts, and insufficient validation of SSH target parameters enables local command execution through malicious flag injection. An attacker can exploit these flaws to achieve arbitrary code execution either remotely via SSH or locally on the system running OpenClaw.

N8N versions up to 1.118.0 is affected by unrestricted upload of file with dangerous type (CVSS 8.8).

n8n is an open source workflow automation platform. [CVSS 8.1 HIGH]

Stored cross-site scripting in n8n's markdown rendering component allows authenticated users to inject malicious scripts into workflows and sticky notes that execute with session privileges when viewed by other users. An attacker with workflow modification permissions can exploit this to hijack sessions and compromise accounts of users who interact with affected workflows. Versions 1.123.9 and 2.2.1 contain fixes for this vulnerability.

n8n has a command injection vulnerability (CVSS 9.9) allowing authenticated users to execute arbitrary OS commands through workflow definitions.

n8n has a TOCTOU race condition vulnerability (CVSS 9.9) enabling bypass of execution restrictions in workflow processing.

Improper Content Security Policy enforcement in n8n workflow automation allows authenticated users to inject persistent XSS payloads into webhook responses that execute with same-origin privileges when other users access the affected workflows. An attacker with workflow creation/modification permissions could exploit this to hijack sessions and compromise user accounts. The vulnerability affects n8n versions prior to 1.123.2.

n8n workflow automation platform has an authenticated code execution vulnerability (CVSS 9.9) through improper runtime behavior modification, enabling server takeover.

Claude Code versions before 2.0.72 allow authenticated attackers to execute arbitrary commands by exploiting a command parsing defect that bypasses the execution confirmation prompt via malicious find command syntax. An attacker with the ability to inject untrusted content into a Claude Code context can trigger unintended command execution with high impact to confidentiality, integrity, and availability. No patch is currently available for affected deployments.

Claude Code versions prior to 2.0.74 allow authenticated users to write files outside designated directories by exploiting inadequate Bash command validation in ZSH clobber syntax parsing. An attacker with the ability to inject malicious content into a Claude Code context window on a ZSH-based system can bypass file restrictions and achieve unauthorized file writes without triggering user permission prompts. This vulnerability requires user interaction and ZSH environment configuration, making it suitable for supply chain or prompt injection attacks against Claude Code users.

Claude Code versions prior to 1.0.111 fail to properly validate trusted domains for WebFetch requests, allowing attackers to register lookalike domains (e.g., modelcontextprotocol.io.example.com) that bypass validation checks. This enables unauthorized automated requests to attacker-controlled servers without user interaction, potentially resulting in sensitive data exfiltration from the user's environment. The vulnerability affects Claude Code's agentic coding functionality and requires upgrading to version 1.0.111 or later to remediate.

Local code injection in NVIDIA Megatron-LM allows authenticated users to achieve arbitrary code execution and privilege escalation through malicious input to vulnerable scripts. An attacker with local access can craft specially designed data to trigger unsafe code evaluation, enabling complete system compromise including data theft and modification. No patch is currently available for this vulnerability affecting all supported platforms.

A flaw was found in Moodle. This cross-site scripting (XSS) vulnerability, caused by improper sanitization of AI prompt responses, allows attackers to inject malicious HTML or script into web pages. [CVSS 7.3 HIGH]

Command injection in OpenClaw's Docker sandbox execution allows authenticated users to manipulate the PATH environment variable and execute arbitrary commands within containers prior to version 2026.1.29. An attacker with valid credentials and ability to control environment variables could achieve code execution within the containerized AI assistant. A patch is available in version 2026.1.29 and later.

Information exposure in vLLM inference engine versions 0.8.3 to before 0.14.1. Invalid image requests to the multimodal endpoint cause sensitive data logging. Patch available.

Amazon SageMaker Python SDK versions prior to v2.256.0 or v3.1.1 disable TLS certificate verification when importing Triton Python models, enabling attackers to perform man-in-the-middle attacks by presenting invalid or self-signed certificates. This vulnerability affects organizations using the affected SDK versions for model imports over HTTPS connections. No patch is currently available for this vulnerability.

Amazon SageMaker Python SDK versions before 3.2.0 and 2.256.0 expose the ModelBuilder HMAC signing key in cleartext API responses, allowing authenticated users with S3 bucket write access to inject malicious artifacts into training jobs that execute with elevated privileges. An attacker with dual permissions to call the DescribeTrainingJob API and modify the training output S3 location can achieve arbitrary code execution when the compromised job runs. No patch is currently available for this vulnerability.

Khoj is a self-hostable artificial intelligence app. Prior to 2.0.0-beta.23, an IDOR in the Notion OAuth callback allows an attacker to hijack any user's Notion integration by manipulating the state parameter. The callback endpoint accepts any user UUID without verifying the OAuth flow was initiated by that user, allowing attackers to replace victims' Notion configurations with their own, resulting in data poisoning and unauthorized access to the victim's Khoj search index. This attack requir...

Unbounded external image fetching in Hugging Face text-generation-inference 3.3.6 VLM mode enables unauthenticated remote attackers to exhaust system resources through malicious Markdown image links in API requests. The router fetches and loads entire image files into memory during input validation regardless of subsequent token limit rejections, causing memory inflation, bandwidth saturation, and potential denial of service on unprotected deployments. This vulnerability affects default configurations lacking authentication and memory limits, with no patch currently available.

A vulnerability in danny-avila/librechat allows attackers to exploit the unrestricted Fork Function in `/api/convos/fork` to fork numerous contents rapidly. [CVSS 5.7 MEDIUM]

The `SimpleDirectoryReader` component in `llama_index.core` version 0.12.23 suffers from uncontrolled memory consumption due to a resource management flaw. [CVSS 5.3 MEDIUM]

In mlflow version 2.20.3, the temporary directory used for creating Python virtual environments is assigned insecure world-writable permissions (0o777). [CVSS 7.0 HIGH]

Arbitrary file write in H2O-3 machine learning platform version 3.46.0.1 allows remote attackers to write data to any file on the server.

In lunary-ai/lunary version 1.2.2, an account hijacking vulnerability exists due to a password reset token leak. A user with a 'viewer' role can exploit this vulnerability to hijack another user's account by obtaining the password reset token. [CVSS 8.8 HIGH]