AI & LLM Vulnerabilities

The tract-onnx library (a Rust crate for running neural network models) has a vulnerability where it loads external data files referenced in ONNX models without checking the file paths. A malicious model can use absolute paths (like `/etc/passwd`) or directory traversal sequences (like `../../../../etc/passwd`) in the `location` field to trick tract into reading arbitrary files on the system and exposing their contents in the model's output. This is a path-traversal vulnerability (a type of attack where an attacker manipulates file paths to access files outside the intended directory).

The githubreceiver component in opentelemetry-collector-contrib has a security flaw where it validates the `required_headers` configuration at startup but never actually checks these headers on incoming webhook requests. This means an attacker can send fake data to the webhook endpoint by bypassing the authentication headers that operators thought were protecting it, especially when the `secret` field is left empty (which skips HMAC validation entirely).

The Sentry exporter in opentelemetry-collector-contrib has a path traversal vulnerability (a type of attack where an attacker manipulates file paths to access unintended locations) because it builds Sentry API URLs by directly inserting the service.name attribute, which remote attackers can control, without checking if it's valid. Since the operator's bearer token (a credential that proves the operator's identity) is automatically added to every request, an attacker can craft a malicious service.name to reach privileged Sentry admin and organization endpoints that they shouldn't have access to.

Claude Code had a security flaw where the domain huggingface.co was automatically approved for the WebFetch tool (a feature that lets AI systems fetch data from the internet), allowing attackers to trick Claude into making web requests to attacker-controlled files and secretly stealing data like files or environment variables without user permission. An attacker would need to first inject malicious content into Claude's context (the information it's working with) to exploit this vulnerability.

Open WebUI has a security flaw where authenticated users can access other users' files through the `/api/chat/completions` endpoint. When a user sends an image request with a file ID (instead of a web link) in the `image_url` field, the server reads that file from disk without checking if the user owns it, then converts it to a format the AI can process. An attacker can exploit this by using another user's file ID to make the AI read and describe private files, leaking their content.

vLLM's audio transcription endpoint checks the size of compressed audio files but not the decoded output, allowing a decompression bomb attack. A 25MB OPUS file can expand to about 14.9GB of PCM audio (raw, uncompressed audio data) in memory, letting an attacker crash the server by exhausting RAM with just a few requests.

vLLM's fix for an earlier vulnerability (CVE-2026-22778) was incomplete, leaving five code paths in the Anthropic router and speech-to-text API that still leak memory addresses. When users send malformed image data, the PIL library raises an error message containing a BytesIO object's memory address (like `<_io.BytesIO object at 0x7a95e299e750>`), and these five locations pass the error directly to clients without removing the address, defeating ASLR (address space layout randomization, a security feature that randomizes where code lives in memory).

vLLM's GGUF dequantize kernels (specialized functions for decompressing quantized neural network weights) use a 32-bit integer parameter to specify tensor size, which truncates when processing very large tensors. This causes the output tensor to be only partially filled by the processing kernel, leaving the remaining memory uninitialized. In multi-tenant deployments (where multiple users share GPU hardware), this unfilled memory may retain sensitive data from other users' previous requests, leaking their information.