Compliance · EU AI Act
The obligations landed on 2 August 2026. Evidence is the deliverable.
For Annex III high-risk systems, testing against prompt injection and data leakage is no longer a good practice you can describe in a policy document — it is a requirement you have to be able to show you met. An agent makes that harder than a model does, because the leak usually happens on a channel nobody exports: a tool call, a shared memory write, a log line. AgentLeak produces that evidence as a file.
Where we are in the calendar
2 Feb 2025Prohibited practices (Art. 5) and AI-literacy duties apply. 2 Aug 2025General-purpose AI model obligations, governance and penalties apply. 2 Aug 2026General application — including the Annex III high-risk obligations. ← in force now2 Aug 2027High-risk AI that is a safety component of a regulated product (Annex I). The headline €35M / 7% of worldwide annual turnover applies to the prohibited practices in Article 5. Failing the high-risk requirements — including the risk-management and robustness duties this page is about — sits in the next tier: up to €15M or 3% of worldwide annual turnover, whichever is higher (Art. 99). Supplying incorrect or misleading information to authorities is up to €7.5M or 1%. Both tiers are large enough to matter; quoting the wrong one at a procurement meeting is not.
Article by article
What the regulation asks for, and the artefact that answers it.
Art. 9 — Risk managementA continuous, documented process across the lifecycle, with testing against reasonably foreseeable misuse.Every run is a dated, reproducible test with a deterministic score, so the process leaves an artefact instead of a claim.Art. 15 — Accuracy, robustness, cybersecurityResilience against third parties altering use or behaviour by exploiting vulnerabilities; explicitly names adversarial examples and model evasion.100 prompt-injection exfiltration scenarios and 46 attack classes, run as a suite rather than ad hoc.Art. 12 — Record-keepingAutomatic logging over the system's lifetime, to a degree appropriate to its purpose.Traces are the record. JSON reports carry findings, channels, severity and a digest.Art. 11 + Annex IV — Technical documentationDocumentation sufficient to show the system meets the requirements.Exportable JSON, HTML and Markdown reports with the finding-to-control evidence matrix.Art. 26 — Deployer obligationsDeployers must monitor operation and keep logs under their control.The gate runs in your CI on your machine; nothing leaves the boundary unless you send it.One report, four frameworks
Findings carry their control mappings, so the same run answers an AI Act file, an ISO audit and a security review without being re-run or re-formatted.
EU AI ActArts. 9, 11, 12, 15 and 26, plus Annex IV technical documentation.ISO/IEC 42001AI management system: operational controls and the evidence that they ran.NIST AI RMFMEASURE — documented, repeated testing with a metric that does not move on its own.OWASP LLM Top 10LLM01 prompt injection, LLM02 sensitive information disclosure, LLM06 excessive agency.Also mappedGDPR (incl. arts. 5, 9, 25, 32), HIPAA, PCI DSS v4.0, Quebec Law 25 and 7 more.This is tooling to make a review faster and better evidenced, not legal advice and not a conformity assessment. No report from any tool makes a system compliant; a notified body, your own risk assessment and your documentation do. What we remove is the part where you cannot show what you tested.
Who this is for
The bundled corpus covers healthcare, finance, legal and corporate operations — which is not a coincidence. Annex III high-risk categories cluster in exactly those places: access to essential private services and creditworthiness, employment and worker management, education, and administration of justice.