Get a free audit
Apple credit Apple credits the researcher who runs our AI penetration testing.

Field Notes / Compliance

AI red teaming and the EU AI Act: what Articles 9, 15 and 55 require, now the high-risk deadline has moved to December 2027

The EU AI Act asks for adversarial testing in three places: the risk-management loop for high-risk systems, the cybersecurity duty in Article 15, and the evaluation duty on the largest general-purpose models. Here is who each one catches, the dates that apply after the AI Omnibus, and the evidence a red team hands you.

Author
Red Team Partners
Read
13 MIN READ
Filed
13 March 2026
Updated
13 November 2026

01 The dates after the AI Omnibus

ost teams read the 2024 headlines and set their plans to 2 August 2026. Then the European Parliament and Council adopted the AI Omnibus. It was published as Regulation (EU) 2026/1744 on 24 July 2026 and came into force three days later, six days before the original high-risk date National Law Review .

ObligationApplies fromTesting duty inside it
Prohibited AI practices2 February 2025None; the practice is banned
General-purpose AI models2 August 2025Article 55: documented adversarial testing for models with systemic risk
Stand-alone high-risk systems (Annex III)2 December 2027, moved from 2 August 2026Article 9 testing and Article 15 resilience
High-risk safety components of regulated products (Annex I)2 August 2028Article 9 testing and Article 15 resilience

Addleshaw Goddard's summary of the adopted text gives the same two dates for high-risk systems Addleshaw Goddard , and so does Loyens & Loeff's note on the amendments adopted under the Digital Omnibus package Loyens & Loeff . The delay does not cover everything. Most transparency duties still apply from 2 August 2026, and generative AI systems already on the market before that date have only until 2 December 2026 to mark their output National Law Review .

The fines stayed where they were. Under Article 99, a breach of the high-risk obligations can cost up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. Prohibited practices reach 35 million euros or 7%. SMEs and start-ups pay the lower of the two figures EU 2024/1689 .

02 Whether your AI counts as high-risk

Article 6 and Annex III list the uses that make a system high-risk. You are probably caught if your AI drives decisions in employment, such as CV screening or performance monitoring; credit and insurance, such as scoring or pricing; critical infrastructure, such as energy, water or transport; education, such as admissions or grading; law enforcement, migration and border control; access to essential public and private services, such as healthcare or benefits; the administration of justice and democratic processes; or biometric identification and categorisation.

The second route catches teams off guard. If your AI acts as a safety component of a product already covered by EU product law, such as a medical device, a vehicle or a machine, it is high-risk through Annex I. You do not choose the label. The use case chooses it for you.

03 Where the Act asks for adversarial testing

You will find the testing duty in three articles. Each asks a different question, and a well-scoped red team answers all three with one body of evidence.

Article 9, risk management. Providers must run a risk-management process across the whole life of a high-risk system. It identifies known and reasonably foreseeable risks, then estimates and evaluates them, including under conditions of reasonably foreseeable misuse. Article 9(6) says the system must be tested to find the right mitigations. Article 9(8) says testing happens during development and in any case before the system reaches the market or goes into service, against metrics and probabilistic thresholds you define in advance. Article 9(5) then asks you to judge the residual risk acceptable EU 2024/1689, Art. 9 .

Attackers have documented prompt injection, system-prompt extraction and retrieval poisoning in public for years, and the OWASP Top 10 for LLM Applications lists them as LLM01, LLM07 and LLM08 OWASP LLM 2025 . A regulator will struggle to see them as unforeseeable. Our notes on system prompt security and RAG, the security blind spot show how each of these attacks works against a live deployment.

Article 15, accuracy, robustness and cybersecurity. Article 15(5) requires high-risk systems to resist unauthorised third parties who try to alter their use, outputs or performance by exploiting weaknesses. It names the AI-specific attacks: poisoning of training data, poisoning of pre-trained components, inputs crafted to make the model err, confidentiality attacks and model flaws. That list reads like the scope of an AI red team.

Article 55, general-purpose models with systemic risk. Providers of the largest models must evaluate them with state-of-the-art protocols, including adversarial testing that they conduct and document. This duty has applied since 2 August 2025. If you build on such a model without training one, the duty sits with your model provider, and your own system still falls under Articles 9 and 15 if it is high-risk.

04 The evidence a red team hands you

Under Article 43, most Annex III providers assess their own conformity through internal control, with no notified body involved. Biometric systems are the main exception. That means nobody signs off your testing for you. Your technical file has to stand on its own when a market surveillance authority reads it.

We map each part of an engagement to the article it answers, so your compliance lead can file it without translation.

What we doWhat you fileArticle it answers
Threat model of the deployed systemThe foreseeable-misuse register9(2)
Attacks across the OWASP LLM Top 10Proven findings with severity and reach9(6), 15(5)
Agreed pass and fail criteria before testingMetrics and thresholds defined in advance9(8)
Fix list and retest of each fixed findingProof the mitigations work9(5), 9(6)
Residual-risk statementWhat remains and why it is acceptable9(5)

A CREST-certified operator verifies every finding before it reaches you. Each one carries its OWASP category, so the same report serves your engineers, your board and your auditors. The researcher who runs our AI penetration testing was credited by Apple on its Web Server Security Acknowledgements for August 2026, for an issue in how AI-generated output was rendered by the application around it. That layer, where the model meets your software, is where many of the findings in an Article 15 test sit.

05 Using the extra time well

You gained sixteen months on Annex III. The teams that use them well test early, while a finding still costs a design change. After launch it can cost a recall. Article 9(8) points the same way: testing belongs inside development, before launch.

Start with the systems that can read untrusted content and then act. Our note on AI red teaming vs penetration testing explains why those need both kinds of test. For a structured pass across every OWASP class, see AI and LLM penetration testing; for the wider governance picture, see our AI security assessment. Once the findings are closed, RTP Robin lets your operators re-run the same attacks as the system changes, which keeps the Article 9 loop current between formal assessments.

Compliance action log
  • List every AI system you place on the EU market and classify each against Annex I and Annex III
  • Write down the pass and fail criteria for each high-risk system before testing starts
  • Commission adversarial testing of the deployed system, including retrieval and tool permissions
  • Fix critical and high findings, then retest to confirm each one is closed
  • Record residual risk in the technical file and set a retest cycle for every material change

To scope the work, book a scoping call with an operator. You leave knowing what the test covers and what you get back.

Questions readers ask

Does the EU AI Act require AI red teaming?

In three places. Article 9 requires providers of high-risk AI systems to test them against risks that include reasonably foreseeable misuse, before the system reaches the market and against metrics defined in advance. Article 15(5) requires high-risk systems to resist attempts by unauthorised third parties to alter their use, outputs or performance, and names data poisoning, model poisoning, adversarial examples and confidentiality attacks. Article 55 requires providers of general-purpose AI models with systemic risk to conduct and document adversarial testing. The Act does not prescribe a red team by name for every system, but a red team is the most direct way to produce the evidence those articles ask for.

When do the EU AI Act high-risk obligations apply now?

The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the dates. Stand-alone high-risk systems listed in Annex III must comply from 2 December 2027. High-risk AI that acts as a safety component of a product under EU product legislation must comply from 2 August 2028. The obligations for general-purpose AI models have applied since 2 August 2025 and were left in place.

Does the EU AI Act apply to companies based outside the EU?

Yes, if you place an AI system on the EU market or its output is used in the EU. A provider in London, Singapore, Manila or New York selling a high-risk system into the EU carries the same obligations as one based in Frankfurt.

What if our AI system runs on a third-party model such as OpenAI, Anthropic or Google?

The model provider carries the general-purpose AI duties. You still carry the duties for the system you built on top of it: how the model is wired in, what data it can read, what tools it can call and what decisions it drives. Most of the attack paths an AI red team finds sit in that layer, the prompts, retrieval and tool permissions you control.

Is internal red teaming enough for the EU AI Act?

The Act does not say an external firm must do the testing. For most Annex III systems, Article 43 lets the provider assess conformity itself under internal control, with no notified body. That puts the weight on your own evidence. An independent tester gives you findings nobody on the build team had a stake in, and a report a market surveillance authority can read without taking your word for it.

What are the fines under the EU AI Act?

Under Article 99, breaching the obligations for high-risk systems can cost up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. Prohibited practices reach 35 million euros or 7%. Supplying incorrect or misleading information to authorities reaches 7.5 million euros or 1%. For SMEs and start-ups the lower of the two figures applies.

References

Sources

  1. European Parliament and Council. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 6, 9, 15, 43, 55 and 99. Text via the AI Act Explorer. artificialintelligenceact.eu
  2. National Law Review. EU Digital Omnibus on AI enters into force (Regulation (EU) 2026/1744, published 24 July 2026, in force 27 July 2026). natlawreview.com
  3. Addleshaw Goddard. EU AI Act: AI Omnibus formally adopted, 1 July 2026. addleshawgoddard.com
  4. OWASP. Top 10 for Large Language Model Applications, 2025 Edition. OWASP Foundation, 2025. genai.owasp.org
  5. NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, 2023. nist.gov
  6. Loyens & Loeff. EU AI Act: key amendments adopted under the Digital Omnibus package. loyensloeff.com