MCP Vulnerability Dataset Preview

This community research artifact publishes ten English records in JSON, JSONL and Parquet, with a native Hugging Face viewer. It distinguishes a product’s use of a language model from whether the documented exploit primitive requires one. Labels come from AI-assisted static review rather than independent human validation or a newly reproduced exploit suite. The repository was created November 4; its documentation says prepared November 3. Original annotations carry CC BY 4.0 where applicable, while CVE material and linked third-party works retain their own terms.