Why this matters now
As of 2 August 2026, the obligations for high-risk AI systems under the EU AI Act (Regulation (EU) 2024/1689) are enforceable. If your agent is used in a high-risk context — recruitment, credit, essential services, and other Annex III areas — or is a general-purpose model provider, a set of concrete technical requirements now applies. Several of them are, in practice, testing and logging requirements: you have to show that the system is accurate and robust against attack, that its operation is logged, and that you monitor it after release.
The Act doesn't name "prompt injection" or "LLM agent." It speaks in terms of accuracy, robustness, cybersecurity, record-keeping, and post-market monitoring. This page translates the three articles that map most directly to agent testing into what you'd actually build and run.
The three articles that map to agent testing
Article 15 — Accuracy, robustness & cybersecurity
Requires resilience to adversarial inputs: data poisoning, adversarial examples, model evasion, confidentiality attacks. For an agent, that's your prompt-injection, jailbreak, and data-exfiltration testing. The article with the most direct testing obligation.
Article 72 — Post-market monitoring
Providers must actively collect and review performance data after deployment. For an agent, that's continuous evaluators running over live traffic — not a one-time pre-launch scan.
Article 12 — Record-keeping (logging)
High-risk systems must automatically log events over their lifetime for traceability. For an agent, that's your trace/span data — the same production traces evsec reads to generate evals. Progress Observability is where that logging lives.
Annex IV — Technical documentation
The file you must be able to produce: a description of the system and the validation/testing done, including metrics and the measures against the risks in Article 15. Your eval definitions and their results are part of this record.
What "evidence" looks like in practice
The gap most teams hit isn't understanding the articles — it's that the articles ask for documented, repeatable testing and a plain checklist doesn't produce any. Evidence here means:
- A defined evaluator for each risk category (injection resilience, PII confidentiality, tool-misuse resistance), with its pass/fail criterion written down.
- Results from running those evaluators over a representative sample of real traffic, with dates and version, not a single spot check.
- A monitoring setup that keeps running post-release, so you can show ongoing conformity (Article 72), not just a launch-day snapshot.
- Traceable logs (Article 12) that tie a finding back to the specific run that produced it.
How evsec produces it
evsec reads a sample of your agent's real traces and generates evaluator packs — LLM-judge prompts plus the span-selection config to run them continuously — and adversarial attack datasets, each tagged to the OWASP LLM Top-10 category and the EU AI Act article it supports. You run them in Progress Observability, which keeps the results and the logs. That chain — real logs → generated evaluators → continuous results → documentation — is the shape of the evidence Articles 15, 72, and 12 ask for.
Generate a full evaluator pack and attack set from a bundled sample trace — tagged to OWASP and the AI Act, with an evidence summary you can download. No signup, no key.
FAQ
Is my AI agent "high-risk" under the Act?
It depends on use, not technology. Agents deployed in Annex III areas (employment, credit, essential private and public services, law enforcement, and others) are typically high-risk; many customer-support or internal-productivity agents are not, though transparency obligations can still apply. Classify with counsel — this page assumes you've concluded the high-risk obligations are in scope.
Does adversarial testing actually satisfy Article 15?
Article 15(5) requires high-risk systems to be resilient against attempts to alter their use or performance by exploiting vulnerabilities, and names adversarial examples and model evasion explicitly. Adversarial testing is how you demonstrate that resilience; it is a core part of the evidence, alongside accuracy metrics and the design measures. See the Article 15 guide.
We're not in the EU. Does this matter?
The Act applies to providers and deployers whose systems' output is used in the EU, regardless of where they're established. If you serve EU users, scope it.