The EU AI Act (Regulation (EU) 2024/1689) is the world’s first binding AI law, and it puts adversarial testing at the centre of compliance. Organisations deploying high-risk AI in EU markets face an August 2, 2026 deadline. Miss it and the penalties reach EUR 35 million or 7% of global turnover, whichever is higher. That makes the AI Act the most financially consequential AI regulation anywhere.

For CISOs and compliance teams, the Act turns AI red teaming from a best practice into a legal requirement. Article 9 mandates risk management systems that include adversarial robustness testing. Article 15 requires resilience against attempts to exploit system vulnerabilities by unauthorised third parties. In practice, both articles point to the same thing: structured red team assessments against your AI systems before the deadline hits. This guide covers the security requirements, maps them to red team activities, and sets out a realistic timeline for compliance.

EU AI Act Overview and Timeline

The EU AI Act was adopted by the European Parliament on March 13, 2024, and entered into force on August 1, 2024. It establishes a risk-based regulatory framework with requirements proportional to the level of risk an AI system poses.

Key Compliance Deadlines

DateMilestoneWhat Takes Effect
August 1, 2024Entry into forceRegulation published, compliance clock starts
February 2, 2025Phase 1 enforcementProhibitions on unacceptable AI practices (Article 5)
August 2, 2025Phase 2 enforcementGPAI model obligations (Articles 51-56), governance structure
August 2, 2026Phase 3 enforcementFull high-risk AI system requirements (Articles 6-49), conformity assessment
August 2, 2027Phase 4 enforcementCertain Annex I AI systems (specific product safety legislation)

The August 2, 2026 deadline is the most significant for security professionals because it activates the full suite of obligations for high-risk AI systems, including the adversarial testing requirements that AI red teaming directly addresses.

Risk Classification

The EU AI Act classifies AI systems into four risk tiers:

Risk LevelDescriptionRegulatory TreatmentExamples
UnacceptableAI systems that pose a clear threat to safety, livelihoods, or rightsProhibitedSocial scoring, real-time biometric identification in public spaces, manipulation of vulnerable persons
High-riskAI systems that significantly impact safety or fundamental rightsStrict obligations (conformity assessment, risk management, transparency, human oversight)Biometric identification, critical infrastructure, education, employment, law enforcement, border control, justice
Limited riskAI systems with specific transparency risksTransparency obligationsChatbots (must disclose AI interaction), emotion recognition, deepfake generation
Minimal riskAI systems that pose minimal riskNo additional obligations (voluntary codes of conduct)AI-enabled video games, spam filters, inventory management

Article 9: Risk Management System — The Core Security Requirement

Article 9 is the most critical provision for AI security professionals. It mandates that providers of high-risk AI systems establish, implement, document, and maintain a risk management system throughout the AI system’s lifecycle.

Article 9 Requirements

Article 9(1): The risk management system must be a continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system, requiring regular systematic updating.

Article 9(2): The risk management system shall comprise the following steps, carried out on a continuous basis:

(a) Identification and analysis of known and reasonably foreseeable risks that the high-risk AI system can pose to health, safety, or fundamental rights

(b) Estimation and evaluation of risks that may emerge when the system is used in accordance with its intended purpose and under conditions of reasonably foreseeable misuse

(c) Evaluation of other risks possibly arising based on analysis of data gathered from post-market monitoring

(d) Adoption of appropriate and targeted risk management measures

Article 9(5): High-risk AI systems shall be tested for the purposes of identifying the most appropriate and targeted risk management measures. Testing shall ensure that high-risk AI systems perform consistently for their intended purpose and that they are in compliance with the requirements set out in this Chapter.

Article 9(6): Testing procedures shall be suitable to achieve the intended purpose of the AI system and do not need to go beyond what is necessary to achieve that purpose.

Article 9(7): Testing shall be carried out at appropriate points throughout the development process, and in any event prior to the placing on the market or putting into service.

Article 9(8): When implementing the risk management system, providers shall give consideration to whether, given its intended purpose, the high-risk AI system is likely to adversely impact persons under the age of 18 or, as appropriate, other vulnerable groups.

How AI Red Teaming Satisfies Article 9

AI red teaming directly satisfies multiple Article 9 requirements:

Article 9 RequirementAI Red Teaming Activity
Identification of known and foreseeable risks (9.2a)OWASP Top 10 for LLMs assessment, threat modeling, attack surface mapping
Risks under conditions of misuse (9.2b)Adversarial testing simulating attacker behavior — prompt injection, jailbreaking, data extraction
Testing for risk management (9.5)Full AI red team assessment covering all vulnerability classes
Testing at appropriate points (9.7)Pre-deployment testing, continuous monitoring, post-update regression testing
Consideration of vulnerable groups (9.8)Bias testing, safety evaluation, harmful content assessment

“The EU AI Act does not use the term ‘red teaming’ explicitly in Article 9, but the requirement for testing under conditions of ‘reasonably foreseeable misuse’ — including adversarial scenarios — is functionally equivalent. Regulators will expect to see evidence of structured adversarial testing.” — Dr. Lilian Edwards, Professor of Law, Innovation and Society, Newcastle University

High-Risk AI Classification (Annex III)

Annex III of the EU AI Act defines the categories of AI systems classified as high-risk. Organizations must determine whether their AI systems fall under these categories, which triggers the full suite of Article 9 obligations including adversarial testing.

Annex III Categories

AreaHigh-Risk AI Applications
1. BiometricsRemote biometric identification, emotion recognition in workplace/education, biometric categorization by sensitive attributes
2. Critical infrastructureAI as safety components in management/operation of critical digital infrastructure, road traffic, water/gas/heating/electricity supply
3. Education and trainingAI determining access to education, evaluating learning outcomes, monitoring prohibited behavior during tests
4. EmploymentAI for recruitment, job advertisements, application filtering, evaluation of candidates, promotion/termination decisions, task allocation, performance monitoring
5. Essential servicesAI for credit scoring, risk assessment in life/health insurance, evaluation of eligibility for public benefits
6. Law enforcementAI for risk assessment of natural persons, polygraphs, evaluation of evidence reliability, profiling in criminal investigations
7. Migration and border controlAI for risk assessment of irregular migration, examination of visa/residence applications, polygraphs
8. Justice and democracyAI for researching and interpreting facts and law, applying law to facts, alternative dispute resolution

Determining If Your AI System Is High-Risk

A practical decision tree:

  1. Is the AI system used in any Annex III area? If yes, it is likely high-risk.
  2. Is the AI system a safety component of a product covered by EU harmonization legislation listed in Annex I? If yes, it is high-risk.
  3. Does the AI system perform profiling of natural persons? If yes and in a listed area, it is high-risk.
  4. Exemption check: Article 6(3) provides an exemption for AI systems that do not pose significant risk of harm to health, safety, or fundamental rights — but this exemption requires documentation and is narrowly construed.

Important note for enterprise AI deployments: Many enterprise LLM applications that are not obviously “high-risk” may still fall under Annex III categories. An AI chatbot used for HR screening is high-risk (employment). An AI system that assists with credit decisions is high-risk (essential services). An AI-powered knowledge management system used in critical infrastructure operations may be high-risk.

GPAI Model Obligations (Articles 51-56)

General Purpose AI (GPAI) models — including foundation models and large language models — have specific obligations under the EU AI Act, effective August 2, 2025.

Obligations for All GPAI Providers

Article 53 requires providers of GPAI models to:

  • Draw up and maintain technical documentation, including training and testing processes and their results
  • Provide information and documentation to downstream providers of AI systems
  • Establish a policy to comply with EU copyright law
  • Publish a sufficiently detailed summary of training data content

Additional Obligations for Systemic Risk GPAI

Article 55 imposes additional obligations on GPAI models with systemic risk (those trained with more than 10^25 FLOPs, or designated by the European Commission based on capabilities):

(a) Perform model evaluation in accordance with standardized protocols and tools, including conducting and documenting adversarial testing of the model

(b) Assess and mitigate possible systemic risks, including their sources

(c) Keep track of, document, and report relevant information about serious incidents and possible corrective measures

(d) Ensure an adequate level of cybersecurity protection for the GPAI model

Article 55 explicitly requires adversarial testing — making AI red teaming a direct legal obligation for providers of frontier AI models. This applies to OpenAI (GPT-4 and successors), Anthropic (Claude), Google (Gemini), Meta (Llama), and any other provider whose models meet the systemic risk threshold.

Implications for Downstream Deployers

Organizations that deploy GPAI models in their applications (as most enterprise AI deployments do) have a shared responsibility:

  • The GPAI provider is responsible for model-level adversarial testing
  • The deployer is responsible for application-level adversarial testing — including prompt injection testing, RAG pipeline security, tool-use exploitation, and system-level access control
  • Both parties must maintain documentation of testing activities and results

Penalty Structure

The EU AI Act establishes a graduated penalty structure that makes non-compliance extremely costly.

Financial Penalties

Violation TypeMaximum Penalty (Enterprises)Maximum Penalty (SMEs/Startups)
Prohibited AI practices (Article 5)EUR 35 million or 7% of global annual turnoverLower caps apply, proportionate to size
High-risk AI obligations (Articles 6-49)EUR 15 million or 3% of global annual turnoverLower caps apply
GPAI model obligations (Articles 51-56)EUR 15 million or 3% of global annual turnoverLower caps apply
Incorrect, incomplete, or misleading information to authoritiesEUR 7.5 million or 1% of global annual turnoverLower caps apply

For context, 7% of global annual turnover for the largest technology companies:

  • Apple (USD 383B revenue): ~EUR 24.5 billion
  • Microsoft (USD 245B revenue): ~EUR 15.7 billion
  • Alphabet (USD 350B revenue): ~EUR 22.4 billion
  • Meta (USD 165B revenue): ~EUR 10.6 billion

These penalties significantly exceed GDPR penalties (4% of global annual turnover, EUR 20 million maximum), signaling the EU’s seriousness about AI regulation.

Enforcement Mechanism

Each EU member state designates national competent authorities to enforce the AI Act. The European AI Office, established within the European Commission, coordinates enforcement across member states and has direct enforcement authority over GPAI model obligations.

Enforcement powers include:

  • Market surveillance and inspections
  • Access to documentation and technical data
  • Orders to withdraw or recall non-compliant AI systems
  • Financial penalties as outlined above
  • Public disclosure of non-compliance

Mandatory Adversarial Testing: What Regulators Expect

While the EU AI Act does not prescribe specific testing methodologies, the combination of requirements in Articles 9, 15, and 55 establishes clear expectations for adversarial testing.

Minimum Expected Testing Activities

Based on regulatory guidance, technical standards in development (particularly by CEN/CENELEC), and preliminary enforcement signals, organisations should expect regulators to require evidence of:

Testing ActivityApplicable RequirementEvidence Required
Threat modelingArticle 9(2)(a-b)Documented threat model specific to AI risks
Prompt injection testingArticle 9(5), Article 15 (robustness)Test results showing resistance to injection attacks
Data leakage testingArticle 9(2)(a), Article 10 (data governance)Evidence that the AI system does not leak training data or sensitive information
Bias and fairness testingArticle 9(8), Article 10(2)(f)Bias evaluation results across protected characteristics
Robustness testingArticle 15 (accuracy, robustness, cybersecurity)Results of adversarial perturbation testing
Safety evaluationArticle 9(2)(a-c)Evidence that the AI system does not produce harmful outputs
Access control testingArticle 15(4) (cybersecurity)Penetration testing results for AI system access controls
Ongoing monitoringArticle 9(1) (continuous process), Article 72 (post-market monitoring)Evidence of continuous testing and monitoring

Documentation Requirements

Article 11 requires detailed technical documentation that must be maintained and made available to authorities. For adversarial testing, this means:

  • Methodology documentation: What testing frameworks were used (OWASP, ATLAS, NIST AI RMF)
  • Scope documentation: What was tested, what was not tested, and justification for exclusions
  • Results documentation: Detailed findings with severity classifications
  • Remediation documentation: What actions were taken to address identified vulnerabilities
  • Retest documentation: Evidence that remediation was effective
  • Continuous monitoring documentation: Ongoing testing results and trend analysis

Practical Compliance Roadmap: Preparing for August 2, 2026

Phase 1: Assessment (Months 1-2)

AI System Inventory:

  • Catalog all AI systems in use or development
  • Classify each system by risk level (prohibited, high-risk, limited, minimal)
  • Identify GPAI model dependencies
  • Map data flows and AI supply chains

Gap Analysis:

  • Evaluate current AI security testing practices against EU AI Act requirements
  • Identify documentation gaps
  • Assess organizational readiness (skills, resources, processes)

Phase 2: Framework Development (Months 2-4)

Risk Management System:

  • Develop or adapt existing risk management processes to incorporate AI-specific risks
  • Establish AI security testing policies and standards
  • Define roles and responsibilities for AI Act compliance
  • Select testing frameworks (OWASP Top 10 for LLMs, MITRE ATLAS, NIST AI RMF)

Testing Program Design:

  • Define testing scope for each high-risk AI system
  • Establish testing frequency (pre-deployment, continuous, triggered by changes)
  • Select tools and methodologies
  • Identify internal capabilities vs. external service requirements

Phase 3: Implementation (Months 4-8)

Initial Testing:

  • Conduct full AI red team assessments for all high-risk AI systems
  • Perform bias and fairness evaluations
  • Test all OWASP Top 10 for LLMs categories
  • Document all findings and remediation actions

Remediation:

  • Address identified vulnerabilities in priority order
  • Implement architectural security controls (least privilege, sandboxing, monitoring)
  • Strengthen prompt injection defenses
  • Improve data governance and access controls

Phase 4: Documentation and Conformity (Months 8-12)

Technical Documentation:

  • Complete all Article 11 technical documentation requirements
  • Document risk management system (Article 9)
  • Document data governance measures (Article 10)
  • Document testing methodologies and results

Conformity Assessment:

  • For Annex III high-risk systems, complete the conformity assessment procedure
  • Obtain CE marking where required
  • Register in the EU AI database

Phase 5: Ongoing Compliance (Post-Deadline)

Continuous Testing:

  • Establish continuous AI security monitoring and testing
  • Conduct regular AI red team assessments (quarterly recommended)
  • Update risk management system based on emerging threats
  • Maintain documentation currency

Post-Market Monitoring:

  • Implement Article 72 post-market monitoring system
  • Track AI system performance and incidents
  • Report serious incidents to authorities
  • Conduct regular reassessments of risk classification

For organisations seeking professional support in meeting EU AI Act compliance requirements through structured AI red teaming, RedTeamPartner.com provides compliance-focused AI security assessments designed to generate the documentation and evidence that EU regulators expect, with particular expertise in high-risk AI classification, adversarial testing methodology, and conformity assessment preparation.

International Regulatory Landscape

The EU AI Act is the most complete AI regulation, but it is part of a global trend toward mandatory AI security testing.

JurisdictionRegulation/FrameworkStatusAI Security Testing Requirement
EUAI ActIn force, phased enforcementMandatory adversarial testing for high-risk AI
USExecutive Order 14110In effect (October 2023)Red teaming for frontier models, voluntary for others
UKAI Safety Institute frameworkOperationalVoluntary red teaming, pre-deployment testing
SingaporeModel AI Governance FrameworkPublishedRecommended adversarial testing
ChinaInterim Measures for Generative AIIn forceRequired safety assessments
JapanAI Guidelines for BusinessPublishedRecommended risk-based testing
CanadaArtificial Intelligence and Data Act (AIDA)ProposedMandatory risk assessments for high-impact AI

Organizations operating globally should prepare for a convergence of AI security testing requirements across jurisdictions. The EU AI Act’s standards are likely to become the de facto global baseline, similar to how GDPR influenced global data protection standards.

Key Takeaways

  1. The EU AI Act mandates adversarial testing for high-risk AI systems through Article 9 (risk management) and Article 55 (GPAI systemic risk), making AI red teaming a legal obligation rather than a best practice.

  2. The August 2, 2026 deadline activates full enforcement of high-risk AI system requirements. Organizations need 12+ months of preparation to achieve compliance.

  3. Penalties reach EUR 35 million or 7% of global annual turnover — exceeding GDPR penalties and representing existential financial risk for non-compliant organisations.

  4. High-risk AI classification (Annex III) covers more applications than many organisations expect — HR screening, credit decisioning, critical infrastructure, education, and law enforcement all trigger full compliance requirements.

  5. GPAI providers must conduct adversarial testing under Article 55, while deployers must test the application layer — creating a shared responsibility model.

  6. Documentation is as important as testing — regulators will expect detailed records of testing methodology, findings, remediation, and ongoing monitoring.

  7. AI red teaming provides the most direct path to satisfying EU AI Act adversarial testing requirements, with findings mapped to recognized frameworks (OWASP, ATLAS, NIST AI RMF).

  8. Global regulatory convergence means that organisations preparing for EU AI Act compliance will be well-positioned for emerging requirements in other jurisdictions.

Sources and References

  • European Parliament and Council. “Regulation (EU) 2024/1689 (Artificial Intelligence Act).” Official Journal of the European Union. August 1, 2024.
  • European Commission. “AI Act Implementation Guidelines.” 2025.
  • European AI Office. “Technical Guidance on High-Risk AI System Classification.” 2025.
  • CEN/CENELEC. “Standardisation Request M/593: AI Standards Development.” 2025.
  • Edwards, Lilian. “The EU AI Act: A Guide for Security Professionals.” Newcastle Law Review. 2025.
  • OWASP. “OWASP Top 10 for Large Language Model Applications, v2.0.” 2025.
  • NIST. “AI Risk Management Framework (AI RMF 1.0).” 2023.
  • MITRE. “ATLAS: Adversarial Threat Landscape for Artificial-Intelligence Systems.” 2025.
  • Alphabet Inc. FY2024 10-K — confirms $350B revenue
  • Meta Platforms FY2024 Results — confirms $164.5B revenue