MITRE ATLAS (Adversarial Threat Landscape for AI Systems)

Ce qu'il fait

Adversarial ML tactics, techniques and real-world case studies — i.e. attacks against AI systems. Contributions are curated into the ATLAS matrix and case study set.

Évaluation franche

Genuinely consequential within AI security — ATLAS IDs are used by vendors, red teams and compliance products. But the bar is a well-evidenced adversarial technique, and a GitHub PR against a MITRE schema is a high-skill submission. Wrong venue for most non-technical safety concerns.

Comment déposer

Contributions flow through the public data repository github.com/mitre-atlas/atlas-data (CONTRIBUTING.md, pull requests against YAML technique and case study files) and via MITRE's ATLAS contact channels. The repo has active open issues and pull requests, so outside PRs are a live route.

Format

Structured YAML records matching the ATLAS schema; case studies follow a fixed template (summary, incident details, procedure mapped to ATLAS technique IDs, references).

Calendrier

Rolling; batched into periodic ATLAS releases.

Ce qui se passe ensuite

MITRE curators review; accepted content appears in a versioned ATLAS release with attribution and is consumed by security tooling.

Ce qu'il accepte

Adversarial ML tactics, techniques and real-world case studies — i.e. attacks against AI systems. Contributions are curated into the ATLAS matrix and case study set.

Ce qu'il n'accepte pas

Model-behaviour or societal-harm concerns with no adversary. ATLAS is a security threat knowledge base modelled on ATT&CK, not a harm registry. Also not a disclosure channel — it documents attacks after the fact.

Géré par

MITRE, with the Center for Threat-Informed Defense (CTID) and industry partners