PHP object injection in October CMS (composer package october/system) allows an attacker-controlled serialized payload stored in a widget session key to be passed to unserialize() without an allowed_classes restriction, enabling instantiation of arbitrary classes and reachable gadget chains — potentially yielding PHP code execution as the web server user. The flaw is reachable only when the niche opt-in cms.safe_mode setting is enabled and a backend user with CMS markup editing access is not fully trusted, so standard production installations with Safe Mode off are not exposed. Public proof-of-concept code exists and vendor patches ship in 3.7.17 and 4.2.x; exploitation is not confirmed in CISA KEV, and the vendor's own CVSS score (3.3) is far below what the stated code-execution impact would normally imply.
Privilege escalation in October CMS 3.x and 4.x lets an authenticated backend user who holds CMS markup editing rights break out of the Twig Safe Mode sandbox and impersonate an existing superuser. The bypass chains two weaknesses: the Laravel session store is handed to Twig templates with unrestricted method access, and raw SQL methods (selectRaw, whereRaw, orderByRaw, joinSub) leak through Eloquent's __call forwarding so a template can run arbitrary queries and then overwrite the admin_auth session key. Publicly available exploit code exists, but exploitation requires cms.safe_mode to be explicitly enabled, a deliberately untrusted markup editor account, and a pre-existing superuser whose session can be forged — a combination that October CMS states does not occur in standard production deployments.
Improper SSH host key validation in andreashappe/cochise through version 0.4.1 allows a network-positioned attacker to perform Man-in-the-Middle attacks against SSH sessions established by the tool. The vulnerable component is the `asyncssh.connect()` call in `src/cochise/ssh_connection.py`, where the SSH Host Key Handler fails to verify server host keys, enabling silent impersonation of target SSH hosts. A public proof-of-concept exists as GitHub issue #13; no vendor patch has been released and the maintainer had not responded to the disclosure at time of reporting.
OS command injection in HackingBuddyGPT up to version 0.5.0 allows a remote low-privileged attacker to execute arbitrary commands on the host running the tool by manipulating input processed by the `ssh_run_command` function in `src/hackingBuddyGPT/extensions/ssh_run_command.py`. The function fails to sanitize shell metacharacters before constructing and executing SSH commands, creating a classic CWE-78 unsanitized-input-to-shell scenario. A publicly available proof-of-concept exploit exists (GitHub issue #150), and the maintainer has not responded to the disclosure; no patch has been released as of this analysis.
DOM-based cross-site scripting in TripleLift's video-bundle.js ad rendering script allows unauthenticated remote attackers to execute arbitrary JavaScript within a publisher's domain by sending crafted postMessage events without origin validation. The attack requires social engineering a victim into visiting an attacker-controlled page, which then fires malicious postMessage payloads at any publisher page running the TripleLift video ad script. Successful exploitation enables session hijacking and unauthorized DOM manipulation in the victim's browser context on the publisher domain. A publicly available proof-of-concept exists; no KEV listing at time of analysis.
Server-side request forgery in netease-youdao LobsterAI affects the `buildBrowserConfig` function within the Browser Network Configuration component, enabling remote attackers to manipulate the application into making forged outbound network requests to attacker-specified targets. Versions 2026.6.15, 2026.8.28, 2026.9.3, and 2026.9.4 are confirmed affected per EUVD enumeration and NVD CPE wildcard coverage. A public proof-of-concept exploit is hosted on GitHub Gist; no vendor patch has been released, and the originating GitHub issue was closed automatically due to inactivity, suggesting no upstream remediation is forthcoming.
Use-after-free in GPAC's `gf_node_unregister` function (scenegraph/base_scenegraph.c) allows an attacker to cause a crash or limited memory corruption by supplying a crafted MP4, BT, or XMT scene file. The vulnerability arises from improper cleanup ordering of proto field default node values during scene graph teardown, with additional related issues in the BT and XMT scene loaders. A public proof-of-concept exploit exists (GitHub issue #3801 and dedicated write-up), though the CVSS 4.0 score of 2.1 and all-Low impact metrics indicate reliable code execution is unlikely - DoS via crash is the primary realistic outcome. No CISA KEV listing exists.
Use-after-free in GPAC's scene graph management allows remote attackers to cause memory corruption via a crafted multimedia scene file. The flaw exists in `gf_node_changed_internal` within `base_scenegraph.c` and is triggered when GPAC's BT/XMT scene loaders fail to properly track node lifecycle, leaving dangling references to freed nodes in `def_nodes` lists and proto field structures. A public proof-of-concept exploit has been disclosed and the vendor has released a fix in tag abi-16.23. No confirmed active exploitation (KEV) is recorded at time of analysis.
Link-following (symlink) handling in NanoClaw's agent-to-agent attachment forwarding lets an authenticated, low-privileged user cause the service to read or modify files outside the intended attachment directory. The flaw is in forwardAttachedFiles() in src/modules/agent-to-agent/agent-route.ts and affects every NanoClaw release from 2.1.0 through 2.1.17; the attack is network-reachable and requires no user interaction. Public exploit code exists and the vendor fix is identified as commit 3f9ed607b7e7a4872747295f75286f1c377d7c33, but real-world risk is currently low (CVSS 4.0 base 2.1, EPSS 0.29%, no CISA KEV listing).
Memory corruption in marcobambini Gravity scripting language versions 0.9.0 through 0.9.7 allows remote attackers with low-privilege access to trigger undefined behavior via crafted JSON input processed by gravity_json.c, potentially causing a crash or exploitable memory state in embedding applications. The root cause is a two-pass JSON parser design flaw where u.object.values is used as a size counter during the first pass but incremented with pointer arithmetic on a null pointer - undefined behavior per the C standard, confirmed to trap under -fsanitize=undefined. A publicly available proof-of-concept exists as GitHub issue #448; no active exploitation has been identified in the wild.
SQL injection in itsourcecode Leave Management System 1.0 lets a low-privileged, authenticated user manipulate the 'ID' argument of /module/company/index.php to run attacker-controlled SQL against the application's database. The flaw is a classic CWE-89 injection reachable over the network with no user interaction, but it requires an existing account on the application (CVSS 4.0 PR:L; assessed CVSS 3.1 PR:L), which is the main barrier to abuse. Publicly available exploit code exists (a proof-of-concept is linked from the VulDB submission and a GitHub issue), so the practical barrier is low for anyone who already holds credentials; however, EPSS is only 0.25% (16th percentile) and there is no CISA KEV listing, so signals converge on LOW real-world priority despite the injection class.
Out-of-bounds read in FFmpeg's convolution video filter (libavfilter/vf_convolution.c, function setup_3x3) affects releases 4.4.0–4.4.6, 5.1.0–5.1.8, 6.1.0–6.1.4, 7.1.0–7.1.3 and 8.0.0–8.0.1, allowing crafted media input to read past allocated buffers and potentially disclose memory or crash the process. Publicly available exploit code exists via the FFmpeg issue tracker, and the fix commit (8970658472d1 / e24b9820b4) has been published, with corrected releases 4.4.7, 5.1.9, 6.1.5, 7.1.4, 8.0.2, 8.1.1 and 9.0. Impact is limited: CVSS 2.1, and exploitation requires the non-default convolution filter to be applied to attacker-influenced input, so this is a low-severity hardening issue rather than an urgent patch-now event.
Authorization bypass via Insecure Direct Object Reference (IDOR) in SourceCodester Inventory Management System 1.0 allows authenticated remote attackers to access arbitrary invoice records by manipulating the ID parameter in invoice.php. The flaw is classified as CWE-639 (Authorization Bypass Through User-Controlled Key), meaning the application trusts a user-supplied identifier to gate access without server-side ownership verification. A publicly available proof-of-concept exploit exists on GitHub; no public KEV or active exploitation has been confirmed at the time of analysis.
Improper authorization in cosmicstack-labs mercury-agent's PermissionManager.checkShellCommand function (versions 1.1.0 through 1.1.13) allows authenticated remote attackers to bypass the shell command permission gate enforced in mercury-agent/src/capabilities/permissions.ts. A public exploit was released via GitHub issue #101, and the vendor has not responded to the coordinated disclosure. The CVSS 4.0 base score is 2.1 (low), reflecting constrained impact due to authentication prerequisites and limited C/I/A blast radius, with no CISA KEV listing confirmed at time of analysis.
SQL injection in itsourcecode Sales and Inventory System 1.0 allows an authenticated remote attacker to manipulate backend database queries via the unsanitized prodcode parameter in /pages/pro_edit1.php. A public proof-of-concept exploit exists on GitHub, lowering the barrier to exploitation for any attacker holding a low-privilege account. The CVSS 4.0 score of 2.1 reflects constrained partial database read, write, and availability impact rather than full compromise, and no active exploitation has been reported.
Server-side request forgery in cosmicstack-labs Mercury Agent (versions 1.1.0 through 1.1.13) lets an attacker who can influence the `path` argument of the GitHub API handler force the agent to issue requests to arbitrary hosts, including internal services reachable from the agent's network position. The flaw is in the githubRequest function of src/utils/github.ts (CWE-918), and publicly available exploit code exists — the reporter's tracking issue (#81) on the project repository is public. The maintainer was informed early through an issue report but has not responded, so no vendor-released fix is confirmed at time of analysis.
Mercury Agent versions through 1.1.13 gate shell execution by validating commands before canonicalizing them, so a caller with low privileges can craft input that passes the raw-string permission check yet resolves to a protected target, disclosing restricted data over the network. The flaw is a validate-before-canonicalize ordering error (CWE-180) in checkShellCommand within src/capabilities/permissions.ts, and publicly available exploit code exists via the project's own issue tracker. The project was notified early but has not responded, so no vendor fix is available, and the CVSS 4.0 base score of 2.1 reflects only a limited confidentiality impact.
Reflected cross-site scripting in itsourcecode Loan Management System 1.0 allows a remote, unauthenticated attacker to inject arbitrary script through the 'page' argument processed by navbar.php, because the value is echoed back into the HTML response without output encoding. Publicly available exploit code exists (a GitHub issue referenced from the CVE record), and the CVSS 4.0 base score is only 2.1 because exploitation requires the victim to interact and the demonstrated impact is limited to integrity (VI:L) with no confidentiality or availability loss and no scope change. It is not reported as actively exploited and no CISA KEV entry is present.
A use-after-free in GPAC's scene-graph handling (gf_sg_script_load path) lets a crafted MP4/VRML scene file corrupt memory when processed by MP4Box or any application embedding libgpac. Publicly available exploit code exists and the flaw is remotely deliverable, though exploitation requires the victim or an ingestion pipeline to actually open the malicious media file (UI:P). A vendor fix is available in release abi-16.23 / commit 9eb40df4448b88d6a6ce3454657c06f47eff0b24, and the CVSS 4.0 base score is only 2.1 because confidentiality, integrity and availability impacts are all rated Low.
Use-after-free
GPAC's MP4Box (through commit f1219cde) crashes with a null pointer dereference in gf_node_list_get_child when it parses a crafted BT or XMT scene description that targets a non-child node field, giving attackers a denial-of-service against anyone who opens the malicious file. Publicly available exploit code exists (linked from the CVE references), and the vendor has already shipped a fix in release abi-16.23 (commit afca1f1181668d85941d51ed1adf647807d5d975), yet CVSS 4.0 scores the issue only 2.1 because impact is limited to availability and exploitation requires the victim to feed the file to GPAC. The CVE record has no KEV listing, so this is a PoC-level, low-severity crash bug rather than an actively exploited threat.
Denial-of-service in Gitlawb openclaude (all versions up to 0.30.0) arises from improper resource handling in the `waitForCallback` function of the xAI OAuth Callback Handler (`src/services/api/xaiOAuthCallback.ts`), where a manipulated Error argument causes availability degradation. The attack is network-reachable but requires user interaction (CVSS 4.0 UI:P), resulting in a low availability impact with no confidentiality or integrity consequences. A publicly available proof-of-concept exists as GitHub issue #2101; the vendor has been notified but has not yet responded to the disclosure.
Stored cross-site scripting in SourceCodester Inventory Management System 1.0 allows an authenticated remote attacker to inject malicious scripts via the Product_Name parameter handled by /api/products_handler.php in the Product Management Module. When other authenticated users subsequently view affected product listings, the injected payload executes in their browsers, enabling session hijacking, credential theft, or UI redressing. A public proof-of-concept exploit is available on GitHub, lowering the bar for exploitation by non-expert attackers.
Cross-site scripting in SourceCodester Inventory Management System 1.0 allows authenticated remote attackers to inject and execute arbitrary client-side scripts via the `/api/vendors_handler.php` Vendor Management endpoint. A public proof-of-concept exploit has been published on GitHub (gist by fhewm98), confirming the issue is reproducible. Exploitation requires low-privilege authentication and victim interaction, limiting blast radius, but successful attacks can hijack sessions, redirect users, or perform unauthorized actions on behalf of the victim.
Reflected or stored cross-site scripting in SourceCodester Inventory Management System 1.0 allows an authenticated remote attacker to inject malicious JavaScript via the Customer_Name parameter in the Customer Management Module's /api/customers_handler.php endpoint. When a second user (such as an administrator) loads the affected customer record, the payload executes in their browser, enabling session hijacking or unauthorized actions on their behalf. A public proof-of-concept exploit is available on GitHub; no active exploitation (KEV) has been confirmed.
OS command injection in MacCMS10 2026.1000.4055 allows an authenticated administrator to execute arbitrary system commands on the underlying server via the Template Handler component by submitting a crafted URL-encoded path parameter. A public proof-of-concept exploit exists, referenced through a GitHub issue report against the project repository. No vendor patch has been released; the project maintainer has not responded to the disclosure.
Remote code execution in Yot CMS 3.3.0 and 3.3.1 is achievable by an authenticated administrator through the Admin Console's unsanitized use of PHP's eval() function in modsys/console/admin.php, where the text POST parameter is passed directly to eval() without validation. A publicly available proof-of-concept on GitHub demonstrates the attack. Despite the technically accurate RCE classification, exploitation is gated behind high-privilege (administrator) authentication, substantially reducing practical risk compared to the tag alone suggests.
OS command injection in D-Link DWR-M921 firmware 1.1.52 exposes the router to remote arbitrary command execution via the WPS configuration endpoint. The `targetAPSsid` parameter passed to `/boafrm/formWsc` is not sanitized before being incorporated into a system-level command within the Boa embedded web server, allowing an authenticated administrator to execute arbitrary OS commands on the underlying embedded Linux system. A public proof-of-concept exploit authored by H3rmesk1t is available on GitHub, lowering the barrier for abuse significantly. No active KEV listing is present at time of analysis.
OS command injection in D-Link DWR-M921 router firmware 1.1.52 permits authenticated remote attackers to execute arbitrary operating system commands by manipulating the sysCmd parameter submitted to the /boafrm/formsysCmd endpoint in the Boa embedded web server. Successful exploitation gives the attacker shell-level control over the device, enabling full compromise of routing, VPN, and wireless functions. A public proof-of-concept exploit has been disclosed on GitHub; no KEV listing or confirmed in-the-wild exploitation has been recorded at time of analysis.
A stack-based buffer overflow exists in GPAC 26.07.0 within the scene-graph DOM event handler gf_sg_dom_event_bubble (src/scenegraph/dom_events.c), reachable through the MP4Box tooling when it parses a crafted scene/BT/XMT media file. Exploitation is local only — the attacker must get a file processed by an affected build — and the disclosed impact is loss of availability rather than code execution or data disclosure. Publicly available exploit code exists (a PoC archive is linked from the advisory), but EPSS is only 0.17% (7th percentile), there is no CISA KEV listing, and the SSVC decision is exploitation=poc, automatable=no, technical impact=partial, so this is a low-priority, local, denial-of-service-class memory-safety bug rather than an urgent remote threat.
Use-after-free in GPAC 26.07.0's MP4Box component allows a local user to crash the application by supplying a crafted MP4 file containing a malformed BIFS scene graph. The flaw resides in gf_node_deactivate_ex() within scenegraph/base_scenegraph.c, where a scenegraph node with a zero instance count could be dereferenced after release because an assertion (gf_assert) was used instead of a guarded early return. A public POC exploit exists; impact is limited strictly to availability (application crash) with no confidentiality or integrity consequence.
Use-after-free in GPAC 26.07.0's MP4Box scenegraph engine allows a local low-privileged user to crash the application by processing a crafted scene file. The vulnerability exists in gf_node_unregister() within scenegraph/base_scenegraph.c, triggered when proto default field nodes (SFNODE/MFNODE typed) are not properly unregistered before scenegraph teardown, leaving dangling pointers subsequently dereferenced. A public proof-of-concept exploit (poc_09_add.zip) is available, though exploitation is limited to local availability impact with no confirmed active exploitation. No special privileges beyond running MP4Box are required.
Null pointer dereference in GNU libredwg 0.13.4's Layer table encoding function allows a local attacker to crash any application built on the library by supplying a DWG file whose LAYER entity has no material handle assigned. The root cause is commit 27118c40 (
Heap-based buffer overflow in jhead up to 3.3 allows a local attacker with low privileges to corrupt process memory by supplying a crafted WebP image with malformed GPS EXIF tags (TAG_GPS_LAT or TAG_GPS_LONG) to the ProcessGpsInfo() function in gpsinfo.c. A proof-of-concept has been publicly released via GitHub, and the project maintainer has not responded to the responsible disclosure report. No vendor-released patch is available at time of analysis.
Denial of service in GPAC MP4Box (up to commit f1219cde) triggers a process abort via an unconditional reachable assertion (`gf_assert(0)`) inside `xmt_parse_element` in `scene_manager/loader_xmt.c` when parsing a crafted XMT scene file. Local, low-privilege users can crash the process by feeding malformed input to the MP4Box command-line tool; no code execution or data leakage is possible. A public proof-of-concept is documented and the vendor has released a fix in version abi-16.23.
Reachable assertion in GPAC MP4Box's scenegraph node unregistration function allows a local attacker to crash the application by supplying a crafted MP4 file. The vulnerability exists in `gf_node_unregister()` within `scenegraph/base_scenegraph.c`, where a hard `gf_assert(pNode->sgprivate->num_instances)` fires when processing scene graph structures with zero instance counts - a state reachable through malformed MPEG-4 BIFS scene data. No confidentiality or integrity impact exists; impact is limited to availability of the MP4Box process. A public proof-of-concept and a vendor-released patch are both available.
GPAC's MP4Box media packager crashes on a crafted MP4 file due to a reachable assertion in stbl_GetSampleInfos() (src/isomedia/stbl_read.c), which fires when a malformed sample-to-chunk table causes firstSampleInCurrentChunk to exceed the current sampleNumber. All GPAC releases up to commit f1219cde are affected. A public exploit writeup exists; however, exploitation is limited to local denial-of-service-crashing the MP4Box process-with no confidentiality or integrity impact. The vendor has released a patch in tag abi-16.23 replacing the bare gf_assert with a graceful error-return path.
Reachable assertion in GPAC's MP4Box tool (up to commit f1219cde) causes process abort when processing a crafted MPEG-4 scene file: `gf_sm_dump_command_list()` in `scene_manager/scene_dump.c` fires `gf_assert(!sdump->current_com_list)` if the command-list pointer is already populated at dump time, crashing MP4Box. A local attacker with standard user privileges who can supply input to the MP4Box binary can trigger this denial of service. No public exploit confirmed in CISA KEV, but a publicly available proof-of-concept exists; vendor-released patch abi-16.23 is available.
Heap buffer over-read in GPAC's MP4Box component allows a local attacker to crash the process by supplying a crafted media file with malformed SVG attributes. The flaw resides in `gf_svg_attributes_copy` (scenegraph/svg_attributes.c) and manifests when MP4Box parses SVG attribute data without adequate bounds enforcement. A publicly available proof-of-concept exists; no active exploitation has been confirmed. The impact is limited to availability - denial of service via process crash - with no confidentiality or integrity consequence.
Null pointer dereference in GPAC MP4Box up to commit f1219cde causes a local denial-of-service when processing crafted VRML/MPEG-4 scenegraph content via the MP4Box tool. The crash originates in scenegraph/vrml_tools.c, specifically in the cyclic traversal flag logic where a node pointer is dereferenced without a NULL guard. A publicly available proof-of-concept exploit exists, and the patch was merged in release abi-16.23. No public exploitation activity (KEV) has been observed; this is a low-severity, locally-exploitable tool crash with no confidentiality or integrity impact.
Out-of-bounds read in jhead's EXIF parser crashes the application when processing a crafted JPEG file. The Get16u() function in exif.c - which reads 16-bit unsigned integers from EXIF data - fails to validate buffer bounds, allowing a malicious JPEG to trigger a read past the end of an allocated buffer. All versions up to and including 3.3 are affected; the vendor has not responded to the coordinated disclosure, and no patch is available. A public proof-of-concept exploit exists, though exploitation is constrained to local access and causes only a low-severity availability impact (crash/DoS), not code execution or data exfiltration.
A local information-disclosure weakness in cosmicstack-labs mercury-agent through version 1.2.0 lets a low-privileged local user manipulate the Shell Permission Manifest so that PermissionManager.checkShellCommand in mercury-agent/src/capabilities/permissions.ts mis-evaluates which shell commands are permitted. Successful abuse exposes sensitive data readable through the agent's shell capability, though integrity and availability are unaffected. Public exploit code exists (linked issue report) and the vendor has not yet responded, but exploitation is limited to hosts already running the agent and requires local access, giving a very low CVSS 4.0 base score of 1.9.
A heap out-of-bounds read/write leading to a buffer overflow exists in the GNU linker (ld), part of GNU Binutils 2.47, inside the elf_x86_64_relocate_section() function of bfd/elf64-x86-64.c, triggered when the linker processes a malformed x86-64 relocation entry whose roff argument is not validated. Exploitation is local — an attacker must supply a crafted ELF object file to a link operation — and yields only limited confidentiality, integrity and availability impact (CVSS 4.0 base 1.9), but publicly available exploit code exists and a heap OOB write/SEGV proof-of-concept has been published. The flaw is fixed in Binutils 2.48 via upstream commit 471130b39c03623ec6d78ece377ff4da3f6bfe7b; it is not listed in CISA KEV, so active in-the-wild exploitation is not confirmed.
A null pointer dereference in the bfd_putl64() function of GNU Binutils 2.47 (bfd/libbfd.c), reachable through the ld linker component, lets a local user crash the tool by feeding it a malformed ELF file. Publicly available exploit code exists (a GitHub write-up demonstrating a SEGV), and the upstream project was notified via a Sourceware bug report but has not responded, so no fixed release has been published. Severity is low (CVSS 4.0 base 1.9) because exploitation requires local access, yields no confidentiality impact, and is limited to integrity/availability loss of the linking process.
GNU Binutils 2.47 contains a heap buffer overflow in cache_bwrite (bfd/cache.c) that is reachable when the GNU linker 'ld' processes a malformed ELF object file with the --gc-sections and -w options. Exploitation is local and low-impact per the vendor vector (CVSS 4.0 base 1.9, AV:L/PR:L, C/I/A:L), but publicly available exploit code exists, confirming the overflow is reproducible against a released version. The issue was reported upstream via Sourceware Bugzilla bug 34442 and, per the vulnerability record, the project has not responded, so no fixed release has been identified.
Denial-of-service in vLLM's Rust-based tiktoken tokenizer allows a local low-privileged attacker to crash the vLLM process by supplying a crafted vocab file containing duplicate token ranks. The TiktokenTokenizer::new() function in the Rust tokenizer backend previously panicked unconditionally when parsing a tiktoken file with two tokens sharing the same rank value, rather than returning a graceful error. A public proof-of-concept exploit is available via GitHub Gist; no patch has been merged as the fixing pull request remains open.
Out-of-bounds read in GPAC 26.07.0's MP4Box tool occurs when a NULL node pointer is passed to gf_node_del() in scenegraph/base_scenegraph.c, causing a process crash when parsing a crafted media or scene-graph file. The fix commit (afca1f1181668d85941d51ed1adf647807d5d975) addresses multiple related input-validation deficiencies across the scene graph, LASeR decoder, MPEG video reframer, and BT scene loader. A public proof-of-concept (poc_19_nstatx.zip) has been released; however, impact is limited to a local availability-only process crash with no confidentiality or integrity consequence, and no active exploitation is recorded in CISA KEV.
Reachable assertion in GPAC MP4Box up to commit f1219cde causes a process abort when a crafted LASeR/MP4 file is processed with MP4Box. The flaw resides in lsr_exec_command_list() in laser/lsr_dec.c, where gf_assert(!codec->bs) fires unconditionally if the codec bitstream handle is already open at command list entry, triggering an abort rather than a graceful error return. A local, low-privileged user with user interaction required can crash the MP4Box process by supplying a malformed input file. A public exploit write-up exists, and the vendor has released a patch in version abi-16.23.
GNU Binutils 2.47 contains a heap out-of-bounds write in the ELF `.eh_frame` handler `_bfd_elf_write_section_eh_frame`, triggered when a malformed object file causes the library to compute a negative size that is then passed to a `memmove` during section rewriting. The attack is local: an attacker must get a crafted ELF file processed by a BFD-based tool (such as `ld`, `objcopy`, or `strip`) on the target system, after which the process can corrupt heap memory and crash. Exploit code is publicly available, though the CVSS 4.0 base score is only 0.9 because the required privileges, user interaction, and limited observed impact keep real-world severity low; there is no CISA KEV listing, and upstream has not responded to the bug report.
Out-of-bounds memory read in Withsecure Atlant's Capricorn scanning engine allows remote unauthenticated attackers to crash the antivirus engine by submitting a crafted document file for processing, resulting in a denial of service. All Atlant deployments running a Capricorn engine build prior to 2025-01-20_02 are affected. No public exploit code has been identified and no active exploitation is recorded; the low CVSS score of 3.7 reflects high attack complexity and a limited, partial availability impact.