EU AI Act (Regulation (EU) 2024/1689) compliance-focused red team testing for high-risk AI systems — Article 9 adversarial testing, Annex III classification, Annex IV technical documentation, conformity assessment, and Notified Body audit preparation. Enforceable since 2 August 2026 with fines up to €35M or 7% global turnover.
Instrucciones de origen · Vista previa de solo lectura
name
eu-ai-act-compliance-redteam
description
EU AI Act (Regulation (EU) 2024/1689) compliance-focused red team testing for high-risk AI systems — Article 9 adversarial testing, Annex III classification, Annex IV technical documentation, conformity assessment, and Notified Body audit preparation. Enforceable since 2 August 2026 with fines up to €35M or 7% global turnover.
{"domain":"ai-compliance","category":"ai","tool_count":8,"guide_count":1,"mitre":"EUAI-Article-9, TA0040-Impact","owasp":"N/A — regulatory framework, not vulnerability class","keywords":["EU-AI-Act","AI-compliance","Article-9","red-team","Annex-III","Annex-IV","conformity-assessment","Notified-Body"],"last_reviewed":"2026-08-09"}
Skill: EU AI Act Compliance Red Team
Supplementary Files:
payloads.md — Article 9 testing commands, Annex IV documentation generators, conformity assessment scripts, Notified Body audit prep check
test-cases.md — 5 structured test cases covering LLM red team, deepfake classification, bias testing, transparency docs, and robustness adversarial
guides/eu-ai-act-article-9-deep-dive.md — Article 9 clause-by-clause analysis with test mapping
Summary
EU AI Act compliance-focused red team skill domain. Validating that high-risk AI systems satisfy Article 9 (data governance, adversarial robustness, logging, transparency, human oversight) and Annex IV (technical documentation) requirements, with conformity assessment evidence and Notified Body audit preparation.
Domain: ai-compliance | Regulation: EU Regulation 2024/1689 | Enforcement: since 2026-08-02
Description
The EU AI Act (Regulation (EU) 2024/1689) is the world's first comprehensive legal framework for artificial intelligence. After a phased implementation beginning February 2025, the high-risk AI system obligations under Articles 8–15 entered enforceable status on 2 August 2026. National competent authorities (the AI Office plus each Member State's regulator) may now initiate audits, demand technical documentation, and impose penalties.
The maximum penalty: €35 million or 7% of global annual turnover, whichever is higher. For mid-cap enterprises the threshold drops to €15M / 3%, but for systemic-risk general-purpose AI models the maximum still applies.
Article 9 specifically mandates that providers of high-risk AI systems conduct adversarial testing (red teaming) to identify and mitigate known and reasonably foreseeable vulnerabilities. This is distinct from technical AI red teaming (covered in ai-safety-redteam-advanced) — Article 9 testing must be documented in Annex IV technical files, with risk treatment decisions, residual risk acceptance, and retest schedules.
This skill covers the compliance perspective: how to map a high-risk AI system to Annex III categories, run red-team tests that satisfy Article 9 acceptance criteria, generate Annex IV technical documentation artifacts, prepare for conformity assessment, and survive a Notified Body audit.
Skill Identity
Aspect
Value
Type
Compliance testing + documentation generation
Distinguishing feature
Regulatory lens applied to technical red team work
This skill produces compliance evidence; adjacent skills produce vulnerability findings
Why this skill exists
Three converging trends made this skill necessary in 2026:
Enforcement began 2026-08-02. Annex III high-risk systems (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice, democratic processes) must comply. Companies that did not prepare Article 9 evidence in 2025-2026 are now exposed to fines.
Annex IV technical documentation is the audit deliverable. Most technical red team tools (Garak, Counterfit, TextAttack) produce findings; Annex IV requires evidence-attached, traceable, retested-on-schedule findings. There is a documentation gap.
Notified Body audits in late 2026 will be the first wave. Companies facing audit need a structured way to organize the artifacts they have (or generate what's missing).
Differentiation from ai-safety-redteam-advanced
Dimension
ai-safety-redteam-advanced
This skill
Goal
Find technical vulnerabilities
Produce regulatory compliance evidence
Output
Vulnerability report, CVSS scores, PoCs
Annex IV technical file, risk register, audit trail
Audience
Security engineers, CISO
Notified Body auditor, AI Office regulator, compliance team
Tools
OWASP LLM Top 10, jailbreaks, prompt injection
Article 9 test matrix, Annex IV template, conformity checklists
Test cadence
Continuous / on-demand
Per-release + per-change + per-retest-window
Both skills use overlapping tools (Garak, Counterfit, TextAttack) but produce different artifacts for different audiences.
Use Cases
Pre-audit readiness assessment for an enterprise deploying a high-risk LLM (e.g., HR resume screening) — generate the Article 9 evidence trail and identify gaps before a Notified Body arrives
Annex III high-risk classification — determine if a new AI feature (e.g., AI-driven loan approval) falls under Annex III scope and which Article 9 sub-clauses apply
Adversarial robustness testing (Article 15) — run documented red team suites against an in-production model, capturing inputs/outputs/metrics for the technical file
Data governance audit (Article 10) — verify training data provenance, bias mitigation, andDatasheet generation completeness
Transparency documentation (Article 13) — generate user-facing notices, instructions for use, and confidence-score disclosures for deployers
Human oversight effectiveness (Article 14) — test that human reviewers can override AI outputs and detect when the AI is wrong; log override rates and decision quality
Conformity assessment preparation — assemble Annex IV technical documentation (system architecture, model card, training data summary, test results, risk register)
Cross-border deployment check — verify that a model trained and approved in one Member State is acceptable for deployment in another (mutual recognition under Article 60)
Core Tools
Tool
Category
Purpose
License
Garak (NVIDIA)
LLM vulnerability scanner
Probe LLMs for OWASP LLM Top 10 + EU AI Act Article 15 robustness; CSV/JSON output for Annex IV
Apache 2.0
Microsoft Counterfit
AI red-team orchestration
Run adversarial suites against ML models (vision, NLP, tabular); task-level evidence collection
MIT
TextAttack
NLP adversarial
Generate and run adversarial text perturbations for robustness testing
Quantify demographic disparity in model outcomes; supports Article 10 data governance
MIT / MIT
Model Card Toolkit (Google)
Documentation generation
Generate Model Cards (Mitchell et al. 2019) that map to Annex IV §2(c)
Apache 2.0
Datasheets for Datasets templates
Documentation generation
Fill Gebru et al. 2018 datasheet template → Annex IV §2(a)
CC-BY
NIST AI RMF (framework)
Risk management alignment
Map EU AI Act evidence to NIST AI RMF Govern/Map/Measure/Manage for cross-jurisdiction
Public domain
Methodology
Phase 1: Classification & Scope
Determine whether the AI system falls under EU AI Act high-risk (Annex III) or is otherwise regulated (e.g., GPAI with systemic risk under Article 51).
Submit Annex IV file + internal control documentation for conformity assessment (internal control for most Annex III systems, Notified Body for some). Establish post-market monitoring loop (Article 72).
Defense Perspective
Defense Layer
Control
Key Points
Classification (Phase 1)
Document Annex III reasoning in scope.json; require legal review for borderline cases
Misclassification is the #1 audit failure mode — every Notified Body will challenge the classification first
Risk assessment (Phase 2)
Use both top-down (scenario-based) and bottom-up (component-based) risk ID; cross-walk to NIST AI RMF
Article 9 doesn't prescribe a methodology; choosing a recognized framework (NIST AI RMF, ISO/IEC 23894) reduces audit pushback
Test execution (Phase 3)
Reproducibility: pin model commit, random seeds, dependency versions; store raw outputs alongside processed results
Auditors request re-runs; non-reproducible test = invalid evidence
Documentation (Phase 4)
Annex IV file should be machine-readable (JSON) AND human-readable (Markdown); generate both from same source
Auditors use both formats depending on whether they are technical or legal
Conformity (Phase 5)
Internal-control conformity (Annex III default) needs 4 documents: technical file, quality management system, EU declaration of conformity, CE mark registration; Notified Body conformity (Article 6(3)) needs 7 documents
Don't conflate the two conformity paths — the documentation burden differs
Post-market (Art. 72)
Logging requirements under Article 12 + Annex IV §2(f) require automatic capture of training runs, post-deployment events, and serious incidents; serious incidents must be reported to the AI Office within 15 days (Article 73)
Logging must be designed, not bolted on; retrofit is painful
Penalty mitigation
Maintain evidence that provider exercised "due diligence" — Article 9 evidence + corrective action on past findings = strongest mitigation argument
The AI Office has discretion; documented good faith is the strongest defense
Detection Methods
Detection of non-compliance with the EU AI Act — the regulator's perspective.
Regulator Detection (What auditors look for)
Absence of risk register — first thing an auditor asks for; missing = immediate red flag
Test evidence older than 12 months — Article 9 implies currency; stale tests = non-compliance
Annex IV file without traceability — claims without reproducibility (no commit hash, no seed) = invalid
Inadequate logging — Article 12 + Annex IV §2(f) define specific event types; missing any = finding
No post-market monitoring plan — Article 72 requires a documented process; absence = non-compliance
Provider Self-Detection
# Compliance self-check
python3 skills/eu-ai-act-compliance-redteam/scripts/self_check.py --model my_model
# Sigma rule for missing Annex IV artifacts (run in CI)# Detects: PR that trains a model without Annex IV file update
title: Model trained without Annex IV update
logsource:
product: ci
detection:
selection:
event_type: model_train
annex_iv_changed: false
condition: selection
Defense Evasion Techniques
How a non-compliant provider might evade detection (for awareness, not endorsement):
Cosmetic Annex IV — generate a technically-complete Annex IV file that makes claims unsupported by actual test evidence; auditors catch this by requesting raw evidence
Outdated test reports — present last year's tests as current; mitigated by date checks + re-test requests
"Continuous improvement" misdirection — claim the model is continuously improved so individual version tests don't apply; AI Office guidance (2026-Q1) explicitly rejected this argument
GPAI classification avoidance — claim a model with 10^25 FLOPs is below the systemic-risk threshold; auditors will demand training records
Borderline Annex III misclassification — claim a hiring LLM is "assistance" not "filtering" (Annex III IV(a)); regulators have published 12 precedent rulings in 2026-Q2 narrowing this
Transparency washing — user-facing notices buried in 50-page ToS; Article 50 requires "clear and conspicuous" disclosure
Practical Steps
Detailed payloads in payloads.md, complete test checklist in test-cases.md.
Step 1: Classify the AI system
Run classify.py (Phase 1) to determine Annex III scope. If borderline, escalate to legal review.
Step 2: Build the risk register
Use the risk_register_template.json; populate top-down (scenarios) and bottom-up (components).
Step 3: Execute the test plan
For LLMs: Garak + TextAttack. For vision: Counterfit. For tabular: Aequitas + Audit-ML.
Step 4: Generate Annex IV
Run annex_iv.py (Phase 4); review the generated Markdown; sign off with risk owner.
Step 5: Submit for conformity assessment
Internal control (default): self-attest with EU declaration of conformity + CE mark. Notified Body (Article 6(3)): engage auditor.
"Check-box" red teaming — running Garak once and calling it Article 9 compliance. Auditors reject this — testing must be per-release and per-change, with documented residual risk acceptance
Conflating "high-risk" with "important" — only Annex III categories count. An "important" AI for the business that isn't in Annex III is unregulated
GPAI threshold confusion — the 10^25 FLOPs systemic-risk threshold (Article 51) is different from high-risk (Annex III). A model can be both, neither, or one
Neglecting logging design — bolting on Article 12 logging after deployment is 10x more expensive than designing it in
Missing 15-day serious-incident window — Article 73 clock starts when provider becomes aware, not when incident is confirmed. Establish a triage SLA
Cross-border mutual recognition assumed — Article 60 mutual recognition exists but Member State regulators have challenged it in practice for politically sensitive deployments (e.g., law enforcement)
Transparency doc as afterthought — Article 13 transparency is a deployer obligation, not just a provider one. Deployers (companies using the AI) are separately liable
Cross-Reference to Related Skills
ai-safety-redteam-advanced — technical AI red team (OWASP LLM Top 10)
This skill codifies EU AI Act compliance red team practice as of 2026-08. Regulation continues to evolve through EU implementing acts (expected late 2026) and Member State national law. Always verify against the latest EUR-Lex publication before relying on this skill for legal compliance.
