Skip to content

Instantly share code, notes, and snippets.

Show Gist options
  • Select an option

  • Save solar-flare99/a84d030087a5619f31c8e458fea835f3 to your computer and use it in GitHub Desktop.

Select an option

Save solar-flare99/a84d030087a5619f31c8e458fea835f3 to your computer and use it in GitHub Desktop.
EU AI Act compliance requirements for agentic AI systems: timelines, Article 72, GPAI Code of Practice, and runtime governance

EU AI Act: Agentic AI Compliance Requirements for 2025–2026

The EU AI Act is an operative enforcement instrument with live penalty regimes, not a proposal — first-wave GPAI provider obligations took effect August 2, 2025, with high-risk system requirements following in August 2026.

Reference: Prismor Agent Governance | PrismorSec/prismor


What the EU AI Act Requires

The GPAI Code of Practice (General Purpose AI) mandates three things from providers today:

  • Public red-team reports — documented adversarial testing of the model or system before deployment
  • Signed usage logs — tamper-evident records of system activity sufficient for forensic reconstruction
  • Live monitoring plans — documented processes for ongoing behavioral surveillance, not periodic audit

Article 72 extends this further: providers must implement active, systematic monitoring throughout the system's lifetime, not just at deployment. Pre-deployment certification alone does not satisfy Article 72.


Timeline: What's Live vs. What's Coming

Obligation Effective Date Scope
GPAI Code of Practice (provider obligations) August 2, 2025 Foundation model providers
High-risk system requirements August 2026 Systems in Annex III categories
Article 72 post-market monitoring August 2026 All high-risk AI systems

High-risk categories under Annex III include: biometric identification, critical infrastructure management, employment/worker management, access to essential services, law enforcement, migration/border control, and administration of justice.


Incident Reporting Timelines Across the Regulatory Stack

EU AI Act does not operate in isolation. An agentic system acting in a regulated domain inherits the obligations of that domain immediately:

Regulation Requirement Timeline
DORA (financial entities) Initial notification after classifying ICT incident as major 4 hours after classification; classification due within 24 hours of awareness
NIS2 Early warning 24 hours
NY RAISE Act Safety incident reporting 72 hours; $1M first violation
CA SB 53 Safety incident reporting 15 days; $1M per violation
CO SB 24-205 Consumer algorithmic decision-making Up to $20,000 per violation per consumer

The math on manual triage: a DORA-compliant incident classification in 24 hours requires automated detection. An agent executing 10,000 tool calls per hour, with humans reviewing at 50 actions per hour, covers 0.5% of decisions manually.


Domain-Specific Triggers

When an AI agent acts in a regulated domain, it inherits that domain's obligations the moment it acts:

  • Loan denial → GDPR Article 22 + Colorado SB 24-205
  • Medical triage → FDA Software as a Medical Device (SaMD) rules
  • Energy infrastructure management → NIS2 + DORA (for financial entities)

Five Technical Capabilities Required for Compliance

From the regulatory requirements, five concrete technical capabilities are needed:

  1. Risk-tier classification for every model before launch
  2. Ethical review checkpoints integrated into the development lifecycle, not bolted on after
  3. Immutable and signed audit logs of agent actions sufficient for forensic reconstruction
  4. Adversarial testing covering the full OWASP Agentic AI Top 10 attack surface
  5. A kill switch that operates at agent speed (automated enforce mode), not committee speed

From Static Compliance to Runtime Governance

The four capabilities most organizations lack today:

  • Real-time behavioral monitoring that detects when actions diverge from approved workflow baselines
  • Consequence-aware authorization that evaluates what an agent is doing, not what permission level it inherited from its human operator
  • Automated incident classification fast enough to meet DORA's 24-hour classification deadline
  • Trajectory-level explainability sufficient for post-market monitoring under Article 72

Pre-deployment certification loses its value the moment an agent begins accumulating context, loading tools dynamically, or modifying its own configuration.


Cyber Insurance Landscape

Traditional insurers are actively excluding AI liability as of 2026:

  • Verisk/ISO CGL exclusions — effective January 2026
  • WR Berkley — absolute AI exclusion across D&O and E&O products
  • AIG and Great American — filed for similar exclusions

A dedicated AI insurance market is forming (Armilla AI, Testudo, HSB/Munich Re), but each carrier requires demonstrated governance as an underwriting prerequisite. Security posture now directly determines insurability.


US Regulatory Volatility

Executive Order 14365 pushes federal preemption of state AI laws while the Commerce Department evaluates targets. Colorado, California, New York, Texas, Illinois, and New Jersey have each enacted AI-specific legislation with different scopes, conflicting definitions, and different penalty structures. The preemption fight is unresolved.

The only defensible posture: plan for the most restrictive applicable timeline until the landscape settles.


Runtime Enforcement: How Prismor Maps to EU AI Act

Prismor is a Python-native agent governance layer (pip install prismor) that maps directly to the five required capabilities:

EU AI Act Requirement Prismor Capability
Signed usage logs (GPAI CoP) Ed25519-signed audit receipts on every tool call evaluation
Active monitoring (Article 72) Observe mode creates behavioral baseline; enforce mode is the kill switch
Adversarial testing coverage 27/27 controls across OWASP LLM Top 10, OWASP Agentic, NIST AI RMF, EU AI Act
Incident classification speed 0.8 ms per tool call evaluation across 10,000 simulated sessions
Compliance evidence prismor attest generates an Ed25519-signed JSON bundle (posture + inventory + audit head + compliance coverage)

Compliance coverage is data-driven: one YAML checklist pack per framework under prismor/runtime/checklists/ + crosswalk.v1.yaml. New frameworks are additive — no new enforcement mechanism required.

pip install prismor

# Generate a compliance attestation bundle
prismor attest

# View EU AI Act coverage
prismor attest coverage --framework eu-ai-act

Keywords

EU AI Act GPAI Code of Practice agentic AI compliance AI regulation 2025 AI regulation 2026 Article 72 monitoring high-risk AI systems AI audit log signed audit trail agent governance LLM compliance AI incident reporting DORA AI NIS2 AI OWASP Agentic AI NIST AI RMF enterprise AI security AI policy enforcement runtime AI governance how to comply with EU AI Act AI agent audit trail agentic AI risk management

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment