Promptfoo Framework Eu AI Act

EU AI Act

The EU AI Act is the world's first comprehensive legal framework for AI. It uses a risk-based approach: Unacceptable Risk (prohibited, Article 5), High Risk (Annex III), Limited Risk (transparency), Minimal Risk. Promptfoo tests for prohibited practices and high-risk system requirements. Penalties up to €35M or 7% of global turnover.

Use this skill when

• Testing EU AI Act compliance for AI systems
• Testing prohibited practices (Article 5): subliminal manipulation, vulnerability exploitation, social scoring, biometric categorisation, remote biometric ID
• Testing high-risk AI systems (Annex III): biometric ID, critical infrastructure, education, employment, essential services, law enforcement, migration, justice/democracy
• Demonstrating due diligence for EU AI Act compliance

Do not use this skill when

• Testing GDPR → use promptfoo-framework-gdpr (applies alongside AI Act)
• Testing ISO 42001 → use promptfoo-framework-iso-42001
• Testing NIST AI RMF → use promptfoo-framework-nist-ai-rmf
• Testing general LLM app security → use promptfoo-redteam-llm

🛡️ Edge cases (mandatory handling)

• Severe penalties — prohibited practices: up to €35M or 7% of global turnover; high-risk violations: up to €15M or 3%; incorrect info to authorities: up to €7.5M or 1%.
• Phased implementation — 6 months (prohibited practices), 12 months (GPAI rules), 24 months (high-risk Annex III), 36 months (full application).
• Compliance beyond testing — documentation, transparency obligations, human oversight, quality management, post-market monitoring, incident reporting.
• Biometric ID testing requires customization — set redteam.purpose to explicitly state the biometric ID behaviors to test.
• Combine with GDPR — data protection requirements apply alongside the AI Act.

🎯 Core workflow

1. Comprehensive testing — redteam: plugins: [eu:ai-act] covers all prohibited + high-risk.
2. Article 5 prohibited practices — load references/article-5.md for subliminal manipulation, vulnerability exploitation, social scoring, biometric categorisation, remote biometric ID.
3. Annex III high-risk — load references/annex-3.md for biometric ID, critical infrastructure, education, employment, essential services, law enforcement, migration, justice.
4. Compliance requirements — load references/compliance.md for documentation, transparency, human oversight, penalties, timeline.
5. Checklist — run references/checklist.md.

🎯 Core principles (summary)

• 4 risk levels: Unacceptable (prohibited), High (Annex III), Limited (transparency), Minimal (no requirements).
• eu:ai-act — comprehensive plugin; eu:ai-act:art5:* for prohibited practices, eu:ai-act:annex3:* for high-risk systems.
• Article 5 prohibitions: subliminal manipulation, vulnerability exploitation, social scoring, biometric categorisation, remote biometric ID (real-time and post).
• Annex III high-risk: 8 categories — biometric ID, critical infrastructure, education, employment, essential services, law enforcement, migration, justice/democracy.
• Penalties up to €35M / 7% turnover for prohibited practices.

References

• references/article-5.md — all 6 prohibited practices with plugin configs
• references/annex-3.md — all 8 high-risk categories with plugin configs
• references/compliance.md — documentation, transparency, human oversight, penalties, timeline, cross-framework
• references/checklist.md — pre-flight, Article 5, Annex III, compliance, post-run checklist