AI Vulnerability Database (AVID)

Ce qu'il fait

Failure modes in general-purpose AI systems — open-weight generative models, closed-API systems, developer tools, and end-to-end AI applications and agents. Records are split into vulnerabilities and reports, organised by a taxonomy library. Rebuilt in 2026 for agentic AI.

Évaluation franche

Modest but real: it is an actively maintained, citable identifier namespace with 2026-dated records, and it is one of the few places a model-behaviour flaw can get a permanent public ID. Small organisation, small readership; do not expect it to force a vendor response on its own.

Comment déposer

Public 'Submit an AI Vulnerability!' link on avidml.org, which routes to a Google Form. Alternatively contribute structured records via the avidml GitHub organisation. Documentation at docs.avidml.org.

Format

Structured record against AVID's taxonomy (security / ethics / performance dimensions), with reproduction detail. Short, technical, evidence-led.

Calendrier

Rolling.

Ce qui se passe ensuite

Accepted records get an AVID identifier (e.g. AVID-2026-R1407) and are published in the searchable database.

Ce qu'il accepte

Failure modes in general-purpose AI systems — open-weight generative models, closed-API systems, developer tools, and end-to-end AI applications and agents. Records are split into vulnerabilities and reports, organised by a taxonomy library. Rebuilt in 2026 for agentic AI.

Ce qu'il n'accepte pas

Not explicitly documented. In practice it is oriented to reproducible model/system failure modes rather than narrative harm accounts — those belong in AIID.

Géré par

AI Risk and Vulnerability Alliance (ARVA), a nonprofit