Timeline after the Digital Omnibus#
The AI Act entered into force on 1 August 2024 [1]. Obligations phase in. In 2026 the EU adopted the "Digital Omnibus on AI", which postponed the high-risk obligations because harmonised standards and national authorities weren't ready on schedule [4][5]. As reported in the public trackers we cite:
| Date | What applies |
|---|---|
| 2 February 2025 | Prohibited practices; AI literacy obligations |
| 2 August 2025 | Obligations for general-purpose AI (GPAI) model providers, including Article 55 for models with systemic risk |
| 2 August 2026 | Most Article 50 transparency obligations (largely unaffected by the Omnibus) |
| 2 December 2027 | High-risk systems listed in Annex III (e.g. biometrics, employment, credit scoring, education, critical infrastructure), postponed by the Omnibus |
| 2 August 2028 | High-risk AI embedded in products covered by Annex I sectoral legislation, postponed by the Omnibus |
The delay doesn't change what is required, only when. Conformity assessment for a high-risk system requires a track record of testing, which takes time to build, so the 2027 date is closer than it looks for any system launching now.
Article 9: risk management that includes testing#
High-risk systems need a risk-management system that runs across the lifecycle: identify and analyse known and foreseeable risks, estimate and evaluate them, and adopt measures. Testing is part of it. Systems must be tested to find the most appropriate risk-management measures and to check consistent performance for their intended purpose, against prior defined metrics and probabilistic thresholds. That phrase is the regulator asking for exactly what a release gate is: a metric, a threshold and a statistical decision rule decided in advance. Our sample-size note shows how to set those thresholds so they mean something.
Article 15: accuracy, robustness, cybersecurity#
Article 15 [2] requires high-risk systems to achieve an appropriate level of accuracy, robustness and cybersecurity and to perform consistently throughout their lifecycle. Three details matter for test design:
- Declared metrics. Accuracy levels and the relevant metrics must be declared in the instructions for use. Declare them with intervals and a dataset definition, or you're declaring a number you can't defend.
- Robustness to errors and inconsistencies, including from interaction with people and other systems. For agents that means multi-turn, tool-error and adversarial-user scenarios, not just clean inputs.
- Cybersecurity against AI-specific attacks. The article names data poisoning, model poisoning, adversarial examples or model evasion, confidentiality attacks and model flaws. For LLM systems this maps onto prompt injection, poisoning and data-leakage testing; see the OWASP LLM Top 10 mapping.
Article 55: general-purpose models with systemic risk#
Providers of GPAI models classed as having systemic risk must, among other things, perform model evaluation using standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing; assess and mitigate systemic risks; track and report serious incidents; and ensure adequate cybersecurity [3]. Downstream deployers aren't bound by Article 55, but in practice they inherit requests for evidence from their own customers, and their high-risk obligations under Article 15 still apply to the systems they build on top.
Evidence checklist for a high-risk system#
- Risk register linking each foreseeable risk to eval categories and controls.
- Pre-defined metrics, thresholds and decision rules for release, with sample-size justification.
- Versioned eval datasets with provenance and representativeness notes.
- Adversarial test reports with severity, reproduction rates and retest results.
- Runtime logging sufficient for traceability (Article 12) and human-oversight design (Article 14).
- Post-market monitoring: online evaluation, drift alerts, incident handling.
Frequently asked questions#
When do EU AI Act high-risk obligations apply?
Following the 2026 Digital Omnibus on AI, high-risk systems listed in Annex III must comply by 2 December 2027 and high-risk AI in products covered by Annex I by 2 August 2028, as reported by public trackers. GPAI obligations have applied since 2 August 2025.
What testing does the EU AI Act require?
For high-risk systems: testing against pre-defined metrics and probabilistic thresholds as part of risk management (Article 9), and appropriate accuracy, robustness and cybersecurity including resilience to adversarial manipulation (Article 15). For GPAI models with systemic risk: model evaluation including documented adversarial testing (Article 55).
Sources#
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the European Union, 12 July 2024. eur-lex.europa.eu
- EU AI Act, Article 15: Accuracy, robustness and cybersecurity (AI Act Service Desk, European Commission). ai-act-service-desk.ec.europa.eu
- EU AI Act, Article 55: Obligations for providers of general-purpose AI models with systemic risk. artificialintelligenceact.eu
- Digital Omnibus on AI: amendments to the AI Act timeline (overview and text tracker). artificialintelligenceact.eu
- Gibson Dunn. EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines and Other Key Changes. gibsondunn.com