Security vulnerabilities, privacy incidents, safety concerns, and policy updates affecting LLMs and AI agents.
TanStack contains a vulnerability that allowed attackers to publish malicious versions of the software to npm (a package repository where developers download code libraries) under the trusted TanStack identity, potentially distributing credential-stealing malware (software that steals login information). This vulnerability is currently being actively exploited by attackers.
Fix: Apply mitigations per vendor instructions, follow applicable BOD 22-01 guidance for cloud services, or discontinue use of the product if mitigations are unavailable.
CISA Known Exploited VulnerabilitiesGitLab MCP Server (a tool that lets AI agents interact with GitLab) had a critical security flaw in versions before 0.6.0 where its HTTP transport exposed an unauthenticated endpoint (a service that processes requests without checking who is calling it) to any website, combined with a misconfiguration that made it accessible from all network interfaces instead of just locally. This allowed attackers from anywhere to make changes to GitLab repositories using the server operator's stored credentials.
Vowpal Wabbit, a machine learning system, has a vulnerability in its GitHub workflow file where pull request titles are inserted directly into bash commands without proper protection. An attacker can craft a malicious pull request title with shell commands that will execute on the build system before Python runs, since the shell processes the string first. Since pull requests can be opened on any branch without special permission, anyone can trigger this attack.
NVIDIA Transformers4Rec for Linux has a vulnerability where attackers can exploit improper deserialization (unsafe processing of data that was converted into a storable format) of untrusted data, potentially leading to code execution (running commands on the system), data tampering, and information disclosure. The vulnerability is tracked as CVE-2026-24162 and a CVSS score (a 0-10 rating of how severe a vulnerability is) has not yet been assigned by NIST.
A vulnerability (CVE-2026-9540) was found in vllm version 0.19.0 that affects the OpenAI-compatible Serving Path component and can be exploited remotely to cause a denial of service (making a service unavailable by overwhelming it). The vulnerability has a CVSS score (a 0-10 rating of how severe a vulnerability is) of 5.5 (medium severity), and a public exploit is already available.
Kiro CLI, a command-line tool that lets developers use AI to run code and shell commands, has a security flaw (CVE-2026-9255) where it doesn't properly check where input comes from before authorizing tool execution. An attacker on the same computer could trick the tool into running arbitrary commands without the user's permission by sending specially crafted data through stdin (the standard input stream that feeds data into a program).
Network-AI v5.4.4 has a critical authentication bypass where the MCP server (a tool that lets AI models call external functions) defaults to an empty secret, causing all authentication checks to pass unconditionally. Additionally, the server allows requests from any origin (CORS wildcard), so an attacker can trick a user into visiting a malicious website that sends commands to their localhost Network-AI server without needing any password, potentially invoking dangerous tools like config_set and agent_spawn.
Boxlite, a sandbox service for running containers, has a path traversal vulnerability (a security flaw where attackers can access files outside intended boundaries) in how it extracts container images. When processing tar files (compressed archives), Boxlite doesn't validate symlink targets (shortcuts to files or directories), allowing an attacker to create a malicious container image that writes files anywhere on the host system, potentially leading to remote code execution (running unauthorized commands on the computer).
BoxLite is a sandbox service that runs untrusted code in lightweight virtual machines (VMs, which are isolated computing environments). It claims to protect host files by mounting directories in read-only mode (preventing writes), but the vulnerability bypasses this: BoxLite tells the underlying VM system (libkrun) to mount directories without actually enforcing read-only restrictions, and it doesn't limit container capabilities (special permissions), so malicious code can remount directories as read-write and modify files that should be protected.
Pydantic AI had a security flaw where attackers could bypass protections against accessing cloud-metadata endpoints (special internal servers that store sensitive credentials) by encoding the IP address in IPv6 transition forms (IPv4-mapped IPv6, 6to4, or NAT64, which are ways to represent IPv4 addresses using IPv6 format). This flaw only affects applications that explicitly allow local file downloads with the `force_download='allow-local'` setting on URLs that could be influenced by untrusted users.
Twig (a PHP template engine) has a vulnerability where template names in `{% use %}` tags aren't properly escaped, allowing attackers to inject arbitrary PHP code that executes when the template cache loads. This bypasses Twig's security sandbox, giving attackers remote code execution (the ability to run commands on the server).
LiteLLM versions before 1.83.10 have a vulnerability where users can change their own role to proxy_admin (an administrative role) through the /user/update endpoint, giving them full control over the system including all users, teams, and API keys. Even users with org_admin privileges can exploit this flaw without needing to chain it with other attacks.
LiteLLM versions before 1.83.14 have a privilege escalation vulnerability (a security flaw that lets someone gain higher-level permissions than they should have) where authenticated internal users can create API keys (credentials for accessing the system) that grant access to admin-only routes without proper verification. This allows attackers to bypass role-based access controls (the system that restricts what different users can do) and gain full admin privileges.
The `mcp-server-kubernetes` tool had a security flaw where access control settings (environment variables that limit which Kubernetes operations are available) only worked when listing tools, but not when actually running them. This meant an attacker or misconfigured AI agent could bypass these restrictions and run any Kubernetes command, like deleting pods or accessing containers, even if they were supposed to be blocked.
Amazon SageMaker Python SDK has a vulnerability where it stores an HMAC signing key (a cryptographic secret used to verify that model files haven't been tampered with) in plaintext as an environment variable that can be read by anyone with access to certain AWS APIs. An attacker with the right permissions could steal this key, use it to forge valid model files, and run malicious code on the system running the model.
LMDeploy, a model serving tool, hardcodes `trust_remote_code=True` (a setting that allows executing custom Python code from downloaded models) when loading models from HuggingFace. An attacker who can control which model path the system loads could point it to a malicious model repository, causing arbitrary code execution (running any commands they want) with the privileges of the LMDeploy server process. This affects LMDeploy version 0.12.3 and earlier.
MLflow versions up to 3.9.0 have a security flaw in the SearchModelVersions feature (an API endpoint that retrieves information about different versions of machine learning models) that fails to check user permissions properly. This allows any logged-in user to see all model versions and sensitive details across the entire system, which is dangerous in shared environments where different teams should only access their own models.
Flowise has a security flaw in its `/api/v1/chatflows/apikey` endpoint that allows a user with a valid API key to view chatflow configurations (including system prompts, workflow graphs, and credential IDs) from other workspaces, as long as those chatflows don't have an API key assigned. The endpoint returns both the user's own chatflows and all unprotected chatflows across the entire system without filtering by workspace, breaking the isolation between workspaces.
Flowise has a mass assignment vulnerability in its PUT /api/v1/user endpoint that lets authenticated users directly change their password hash without verifying their old password. An attacker with a stolen session token can send a crafted request that overwrites the credential field, bypassing password verification, hashing enforcement, and policy validation, which gives them permanent access to the account.
Flowise, an AI tool, has a hardcoded setting that allows any webpage on the internet to make requests to its text-to-speech (TTS, a feature that converts written text into spoken audio) endpoint using your stored credentials. This bypasses the server's normal cross-origin request protection (CORS, which controls what websites can access a server's data), letting malicious webpages secretly generate speech on your behalf.
Fix: Update GitLab MCP Server to version 0.6.0, which fixes the vulnerability.
NVD/CVE DatabaseFix: This vulnerability is fixed by commit 998e390e80a7e8192d7849b7784bc113dbd190ad.
NVD/CVE DatabaseFix: A pull request to fix this issue awaits acceptance (mentioned in the source as pending at https://github.com/vllm-project/vllm/pull/37594).
NVD/CVE DatabaseFix: Update kiro-cli to version 1.28.0 or later. The affected versions are kiro-cli prior to 1.28.0.
AWS Security BulletinsFix: Upgrade to Pydantic AI version 1.99.0 or later, which extends the blocklists to cover IPv6 transition forms that route to blocked IPv4 endpoints and adds protection for additional IANA-reserved IP ranges. For unpatched versions, avoid using `force_download='allow-local'` on URLs influenced by untrusted input, or resolve hostnames manually and validate them against your own blocklist including IPv6-encoded forms before creating the FileUrl.
GitHub Advisory DatabaseFix: `Compiler::string()` now escapes single quotes in addition to the characters it previously escaped, preventing template names from breaking out of the surrounding PHP string context.
GitHub Advisory DatabaseFix: Update LiteLLM to version 1.83.10 or later.
NVD/CVE DatabaseFix: Update LiteLLM to version 1.83.14 or later.
NVD/CVE DatabaseFix: The fix applies the same filtering logic from the tool listing layer to the tool execution layer in the `CallToolRequestSchema` handler, so that restricted tools return an error when called directly. This was fixed in v3.6.0.
GitHub Advisory DatabaseFix: Upgrade to Amazon SageMaker Python SDK v2.257.2 or v3.8.0. According to the source: 'AWS recommend upgrading to the latest version and rebuilding any models previously created with ModelBuilder using the updated SDK.' As a temporary workaround if upgrading is not immediately possible: 'users can manually remove the SAGEMAKER_SERVE_SECRET_KEY environment variable from existing SageMaker models by recreating the model without this variable in the container environment configuration.'
GitHub Advisory DatabaseFix: The issue is resolved in version 3.10.0.
NVD/CVE DatabaseFix: Remove the hardcoded CORS wildcard headers from the TTS endpoint. Specifically, delete these lines from `packages/server/src/controllers/text-to-speech/index.ts` at line 83: `res.setHeader('Access-Control-Allow-Origin', '*')` and `res.setHeader('Access-Control-Allow-Headers', 'Cache-Control')`. This allows the server's standard CORS middleware to handle access control instead.
GitHub Advisory Database