All stories
MCP Vulnerability Dataset Preview Launches Community reports

MCP security preview separates software flaws from model-dependent attacks

A new ten-record English dataset exposes evidence links and reasoning while distinguishing an application's use of AI from an exploit's need for a model.

By AITrending2 min read
Sources (4)

Why now The public Hugging Face repository was created November 4 at 03:20 UTC; its November 3 preparation date belongs to the documentation, not a newly discovered vulnerability.

An MCP-related security flaw does not automatically require a language model to exploit it. A new community dataset preview makes that distinction an explicit field, alongside the evidence and reasoning behind each classification.

Published by XuViewer on Hugging Face, the MCP Vulnerability Dataset preview contains ten existing CVE records for reviewing its presentation and schema. Repository metadata dates its creation to November 4 at 03:20:41 UTC. The card’s November 3 preparation date describes the documentation; neither date turns the underlying CVEs into newly discovered vulnerabilities.

Follow the evidence behind a label

Readers can browse the native dataset viewer or download JSON, JSONL and Parquet files. The schema separates upstream titles, affected versions and severity information from the project’s own mechanism judgments.

Its model-necessity field asks whether the documented attack primitive or chain requires a language model to participate. A classification that no model is required applies to that described mechanism; it does not rule out an attacker using a model elsewhere.

Other fields identify the untrusted entry, sensitive operation and trust or privilege boundary. Each judgment can point to evidence IDs and a written rationale. An unestablished value remains null, which the schema explicitly says means unknown rather than false.

For security researchers and agent builders, the practical appeal is an inspectable review format: a label can be checked against its stated basis instead of being treated as a verdict simply because a product uses MCP.

A schema preview, not a prevalence study

The card calls these ten records a convenience sample. It does not present them as representative of MCP security, defects in the protocol specification or a new estimate of how often vulnerabilities occur.

The annotations were developed through AI-assisted static review of saved public records and available advisories, code and patches. The authors disclose no independent human validation and do not claim successful runtime reproduction of every exploit. AITrending reviewed the English documentation and schema; it has not independently re-audited the CVEs.

Original contributions are licensed under CC BY 4.0 where the contributors hold rights. CVE-derived material keeps its upstream terms, and links do not relicense third-party reports or source code. The public release is this ten-record preview, not the full private annotation cohort.

Related reading