FLARE-AI (Flaw Reporting for AI)

Ce qu'il fait

A flaw, vulnerability or incident in any AI system. This is the first live implementation of the standardised AI flaw report concept — it covers model-behaviour flaws, biases and incidents, not only security bugs.

Évaluation franche

This is the single most important new mechanism for an unaffiliated person with a model-behaviour finding, because it solves the routing problem: it converts an informal observation into a standard artefact and delivers it to named labs and a national coordination centre at once. Caveats: it launched only weeks before this research, is a research preview, and there is not yet public evidence of outcomes. Backed by CERT/CC's institutional credibility, which is the strongest signal available.

Comment déposer

Guided web form at ai-reports.org ('Report Now'), takes roughly 10 minutes and generates a standardised, machine-readable flaw report. You then choose where to route it: the site lists Anthropic, Hugging Face, Cohere, the UK AI Safety/Security Institute, CERT/CC, ECDA and DAVID as recipients. You can also take the generated report and submit it yourself anywhere, or go direct to CERT/CC's VINCE.

Format

Structured, machine-readable flaw report generated by the form — the standardised schema proposed in the 2025 third-party flaw disclosure work.

Calendrier

Rolling. Went live 1 July 2026.

Ce qui se passe ensuite

The report is transmitted to the recipient(s) you select. Coordination behaviour then depends on the recipient — CERT/CC applies its own coordinated disclosure process.

Ce qu'il accepte

Ce qu'il n'accepte pas

Not explicitly enumerated on the site; it is still labelled a research preview.

Géré par

Carnegie Mellon University Software Engineering Institute (SEI/CERT) with academic and industry partners; open source

Comment les joindre