AI & LLM Vulnerabilities

NVIDIA Triton Inference Server has a vulnerability where it doesn't properly check user inputs and can crash when given a large number of outputs, potentially causing a denial of service (making the server unavailable to users). The vulnerability stems from excessive memory allocation triggered by malformed input.

text-generation-webui, an open-source web interface for running Large Language Models, has a vulnerability in versions before 4.3 where the superbooga and superboogav2 RAG extensions (tools that fetch external documents to help answer questions) accept user-provided URLs without checking them for safety. This allows attackers to access cloud metadata endpoints (services that store sensitive credentials in cloud environments) and steal IAM credentials (identity and access management tokens that control what users can do). The vulnerability is fixed in version 4.3.

text-generation-webui, an open-source web interface for running Large Language Models, has a path traversal vulnerability (a security flaw where an attacker can access files outside the intended directory) in versions before 4.3. An unauthenticated attacker can exploit this by sending specially crafted requests through the API to read any file on the server, because Gradio (the framework it uses) does not validate user input on the server side.

A vulnerability in HuggingFace Transformers' `Trainer` class (a tool for training AI models) allows attackers to run arbitrary code by providing a malicious checkpoint file. The problem occurs because the `_load_rng_state()` method uses `torch.load()` without the `weights_only=True` parameter (a safety setting that restricts what code can run), leaving systems vulnerable when using PyTorch versions below 2.6.

Anthropic's Claude Code CLI and Claude Agent SDK have a vulnerability where attackers can execute arbitrary commands (run any code they want) by inserting shell metacharacters (special characters like $() that tell the system to run commands) into file paths. Even though the code tries to protect these paths by wrapping them in double quotes, the POSIX shell (the command-line interface on Unix/Linux systems) still processes these injected expressions, giving attackers the same permissions as the user running the CLI.

Anthropic's Claude Code CLI and Claude Agent SDK have a vulnerability where attackers can run arbitrary commands by manipulating the TERMINAL environment variable (a setting that controls which terminal program to use). When the software constructs shell commands, it doesn't properly sanitize the TERMINAL variable, allowing attackers to inject shell metacharacters (special characters that have meaning to command interpreters) that get executed with the user's privileges.

KubeAI, a tool that runs AI models on Kubernetes (a system for managing containerized applications), has a vulnerability in versions before 0.23.2 where attackers can inject malicious shell commands (arbitrary code execution instructions) through Model resource creation. The flaw exists because the ollamaStartupProbeScript() function doesn't properly validate user input when building commands that run during startup checks.

Directus has a security flaw in its TUS resumable upload endpoint (a feature that lets users upload files in chunks) that lets any authenticated user overwrite any file in the system by specifying its UUID (unique identifier), bypassing row-level permissions (rules like 'users can only edit their own files'). This can lead to permanent data loss and allow low-privilege users to replace important files with malicious content.

BentoML's Dockerfile generation uses an unsandboxed Jinja2 template engine (a tool that processes template files with dynamic code) with dangerous extensions enabled, allowing attackers to embed malicious code in a template file. When a victim imports a malicious bento archive and runs the containerize command, the attacker's code executes directly on the victim's host machine before any container isolation happens, rather than inside a container where it would be restricted.