EU AI Act and AI red teaming

The EU AI Act requires providers of general-purpose AI models with systemic risk to conduct and document adversarial testing of the model (Article 55), and that obligation has applied since 2025-08-02. For everyone else, no law reviewed for this site names red teaming as a duty, but OWASP, MITRE ATLAS and NIST AI RMF all describe it, and most vendors map their findings to those frameworks.

What Article 55 says

Article 55(1)(a) requires providers of general-purpose AI models with systemic risk to "perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model". The stated aim is to identify and mitigate systemic risks. Article 55 says "adversarial testing", not "red teaming"; vendors sell the same work under the red-teaming name.

Deadlines

• GPAI obligations (Chapter V) have applied since 2025-08-02.
• GPAI models already on the market before that date have until 2027-08-02 to comply.
• Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 2026-07-24 and entered into force on 2026-07-27. It moved Annex III high-risk obligations to 2027-12-02 and Annex I (product-embedded) high-risk obligations to 2028-08-02. The GPAI provisions "continue to apply as planned".

Who this applies to

The Article 55 duty falls on providers of GPAI models with systemic risk, that is, the companies that train and release the models themselves. If you deploy someone else's model in a chatbot or agent, Article 55 does not bind you. Deployers of high-risk systems (Annex I or III) have separate obligations with the later deadlines above; what those obligations require of testing was not reviewed for this site.

Vendors that claim an EU AI Act mapping

11 of the 23 vendors in the comparison claim on their own pages that their reports map to the EU AI Act. Claims are the vendor's; none was verified.

• Bugcrowd
• Cisco AI Defense : EU AI Act appears as "compliance readiness" in a Cisco blog, not as a formal mapping in the data sheet.
• Enkrypt AI
• HackerOne : HackerOne's page cites an "OWASP LLM Top 10 (2026)"; OWASP's project page lists 2025 as the current LLM edition.
• Holistic AI
• Mindgard : Names EU AI Act Article 55 specifically.
• NCC Group
• Noma Security : Offered as presets.
• Pillar Security
• Promptfoo
• Straiker

4 also claim ISO 42001 mappings: Holistic AI, Mindgard, NCC Group, Noma Security. Holistic AI additionally names NYC Local Law 144.

OWASP: the application-level checklist

The OWASP Top 10 for LLM Applications (2025) is the list most vendors map to:

1. LLM01: Prompt Injection
2. LLM02: Sensitive Information Disclosure
3. LLM03: Supply Chain
4. LLM04: Data and Model Poisoning
5. LLM05: Improper Output Handling
6. LLM06: Excessive Agency
7. LLM07: System Prompt Leakage
8. LLM08: Vector and Embedding Weaknesses
9. LLM09: Misinformation
10. LLM10: Unbounded Consumption

OWASP also published the Top 10 for Agentic Applications for 2026 (released 2025-12-09) and a GenAI Red Teaming Guide (released 2025-01-22) that organises red teaming into model evaluation, implementation testing, infrastructure assessment and runtime behaviour analysis. Note: some vendor pages cite an "OWASP LLM Top 10 (2026)"; OWASP's project page lists 2025 as the current LLM edition, and 2026 as the Agentic edition.

MITRE ATLAS: the adversary knowledge base

MITRE ATLAS is "a public knowledge base of adversary TTPs targeting AI systems". The latest data release is 2026.10 (2026-10-28) with 129 techniques, 92 sub-techniques, 44 mitigations and 77 case studies; it added agent-related techniques and case studies. Vendors that map to ATLAS let you express findings as techniques an attacker would actually use.

NIST AI RMF: the US voluntary framework

The NIST AI RMF Generative AI Profile (NIST AI 600-1), published 2024-07-26, defines AI red teaming and names it in several actions:

• MS-2.7-007: "Perform AI red-teaming to assess resilience against" abuse to attack other systems, GAI attacks such as prompt injection, and ML attacks such as data poisoning, membership inference and model extraction.
• MP-5.1-005: "Conduct adversarial role-playing exercises, GAI red-teaming, or chaos testing."
• MS-1.3-003: verify that those conducting structured human feedback exercises "are not directly involved in system development tasks for the same GAI model".

US status as of November 2026: Executive Order 14110 was revoked by the 2025-01-23 order "Removing Barriers to American Leadership in Artificial Intelligence", which also directed an AI Action Plan and the revision of OMB memos M-24-10 and M-24-18. NIST AI 600-1 is still published (page last updated 2026-04-08) and remains voluntary. NIST's AI standards center (CAISI) describes leading unclassified evaluations through voluntary agreements with developers. No source opened for this site shows a binding federal red-teaming mandate for private companies.

Putting it together

• If you provide a GPAI model with systemic risk: Article 55 applies now. You need documented adversarial testing; several vendors (Mindgard names Article 55 explicitly) sell to this requirement.
• If you deploy a high-risk system (Annex I or III): your deadlines moved to 2027-12-02 or 2028-08-02. Article 55's adversarial-testing duty is not yours; the high-risk testing requirements were not reviewed for this site.
• If you deploy an ordinary LLM app: no EU or US law names red teaming for you. Use OWASP (LLM 2025, Agentic 2026), MITRE ATLAS and NIST AI 600-1 as the frameworks your findings map to, and pick vendors that claim those mappings.

Common questions

Does the EU AI Act require AI red teaming?

For providers of general-purpose AI models with systemic risk, yes. Article 55(1)(a) requires them to perform model evaluation including "conducting and documenting adversarial testing of the model". GPAI obligations have applied since 2025-08-02; models already on the market before that date have until 2027-08-02 to comply.

Did the Digital Omnibus change the GPAI red-teaming obligation?

No. Regulation (EU) 2026/1744, in force since 2026-07-27, moved Annex III high-risk obligations to 2027-12-02 and Annex I product-embedded high-risk obligations to 2028-08-02, but the GPAI provisions "continue to apply as planned".

Is AI red teaming legally required in the United States?

No source opened for this site shows a binding US federal red-teaming mandate for private companies. Executive Order 14110 was revoked on 2025-01-23. NIST AI 600-1 remains published and voluntary.

This page summarises public sources; it is not legal advice. Dates and article numbers are as the linked sources stated them on 2026-11-07.