squishai/vex-feed

Community OpenVEX 0.2.0 exploitability statements for ML model families distributed through squish.

Feed URL

https://raw.githubusercontent.com/squishai/vex-feed/main/feed.openvex.json

This is VexCache.DEFAULT_URL — the canonical community feed used by squash when no custom VEX endpoint is configured.

Coverage

CVE	Description	Products	Status
CVE-2024-34359	llama-cpp-python GGUF heap buffer overflow	Qwen2.5, Qwen3, Llama-3.2, Gemma-3	not_affected
CVE-2023-27534	PyTorch pickle deserialization RCE	Qwen2.5, Qwen3, Llama-3.2, Gemma-3, squish@9.13.0	not_affected
CVE-2024-3660	Keras arbitrary code execution via model config	Qwen2.5, Qwen3, Llama-3.2, Gemma-3	not_affected

Rationale

All squish-distributed models are:

• Loaded exclusively via the safetensors library (no pickle / .pt / .bin)
• Served through the MLX backend (no llama-cpp-python, no GGUF)
• Not Keras-format architectures

Contributing

Open a PR to add VEX statements for additional CVEs or models. Each statement must:

1. Reference a published CVE (or well-known advisory ID)
2. Use pkg:huggingface/OWNER/REPO PURLs for model products
3. Include an impact_statement explaining the reasoning
4. Validate against the OpenVEX 0.2.0 schema

License

CC0 1.0 Universal — dedicated to the public domain.