AI & LLM Vulnerabilities

The huggingface/transformers library before version 4.53.0 has a vulnerability where malicious regular expressions (patterns used to match text) in certain settings can cause ReDoS (regular expression denial of service, a type of attack that makes a system use 100% CPU and become unresponsive). An attacker who can control these regex patterns in the AdamWeightDecay optimizer (a tool that helps train machine learning models) can make the system hang and stop working.

Codex CLI (a coding tool from OpenAI that runs on your computer) versions 0.2.0 to 0.38.0 had a sandbox bug that allowed the AI model to trick the system into writing files and running commands outside the intended workspace folder. The sandbox (a restricted area meant to contain the tool's actions) wasn't properly checking where it should allow file access, which bypassed security boundaries, though network restrictions still worked.

Flowise version 3.0.5 contains a Server-Side Request Forgery vulnerability (SSRF, a flaw that lets attackers trick the server into making requests to internal networks on their behalf) in the /api/v1/fetch-links endpoint, allowing attackers to use the Flowise server as a proxy to access and explore internal web services. This vulnerability was patched in version 3.0.6.

A vulnerability in Keras (a machine learning library) allows attackers to run arbitrary code on a system by creating a malicious .keras model file that tricks the load_model function into disabling its safety protections, even when safe_mode is enabled. The attacker does this by embedding a command in the model's configuration file that turns off safe mode, then hiding executable code in a Lambda layer (a Keras feature that can contain custom Python code), allowing the malicious code to run when the model is loaded.

A vulnerability exists in Keras' Model.load_model method where specially crafted .h5 or .hdf5 model files (archive formats that store trained AI models) can execute arbitrary code on a system, even when safe_mode is enabled to prevent this. The attack works by embedding malicious pickled code (serialized Python code) in a Lambda layer, a Keras feature that allows custom Python functions, which bypasses the intended security protection.

CVE-2025-23329 is a vulnerability in NVIDIA Triton Inference Server (a tool used to run AI models efficiently) on Windows and Linux where an attacker could damage data in memory by accessing a shared memory region used by the Python backend, potentially causing the service to crash. The vulnerability involves improper access control (failing to properly restrict who can access certain resources) and out-of-bounds writing (writing data to memory locations it shouldn't).

CVE-2025-23328 is a vulnerability in NVIDIA Triton Inference Server (software that runs AI models on Windows and Linux) where an attacker could send specially crafted input to cause an out-of-bounds write (writing data outside the intended memory location), potentially causing a denial of service (making the service unavailable). The vulnerability has a CVSS score of 4.0, indicating moderate severity.

NVIDIA Triton Inference Server has a vulnerability in its DALI backend (a component that processes data) where improper input validation (the failure to check if data is safe before using it) allows attackers to execute code on the system. The issue is classified as CWE-20, a common weakness type related to input validation problems.

picklescan is a tool that checks if pickle files (a Python format for storing objects) are safe before loading them, but versions up to 0.0.30 have a vulnerability where attackers can bypass these safety checks by giving a malicious pickle file a PyTorch-related file extension. When the tool incorrectly marks this file as safe and it gets loaded, the attacker's malicious code can run on the system.