The EU AI Act conformity-evidence layer for autonomous AI agents.
Case file · 2025 → 2026
Three signals from the last year, each a record that could not be produced or trusted after the fact.
01
Jul 2025Action without a provable record· Fortune · AI Incident Database #1152
An autonomous coding agent deleted a production database holding records on 1,206 executives during an explicit code freeze, then misreported that a rollback was impossible. The developer also reported the agent fabricating records and test results.
02
Jun 2025Indirect prompt injection· CVE-2025-32711
Reported by Aim Labs as the first zero-click exploit found in a widely used generative-AI product, it could exfiltrate sensitive data with no user action. Microsoft patched it server-side and found no evidence of exploitation, leaving only vendor-controlled logs, not tamper-evident proof.
03
Mar 2026No attribution, no evidence· CSA / Aembit survey, n=228
68% of organizations cannot clearly distinguish human activity from AI-agent actions in their own systems.
The blocker is not the AI. It is the lack of verifiable evidence.
docker pull ghcr.io/edoardobambini/iaga-sentinel:v1.8.1No Rust toolchain, demo data pre-seeded.
Three things in one binary. The verdict stays deterministic under fixed risk weights, and reproducible; models produce scores the policy reads, never the decision.
A deterministic risk engine issues every allow, review, or block verdict. Dictum, a typed policy DSL with a Hindley-Milner type checker, layers on top and can only tighten it. Deterministic under fixed risk weights, reproducible, replay-safe.
$iaga policy check no_pii_egress.dictum
Each verdict becomes an Ed25519-signed receipt linked to the previous one in a hash-chained append-log, per run. Replay re-verifies the whole signed chain offline, against one public key.
$iaga replay <run_id> --verify-only
A standalone iaga-verify binary checks every signature and hash-chain link against one public key. No database, no server, no call-home. The proof does not depend on us.
$iaga-verify run.json --key <hex>
The pieces exist separately. We built them as one chain.
The defensible position is the intersection: runtime cryptographic proof, offline-verifiable, and deployable entirely on EU infrastructure you control today, with eIDAS-qualified signatures and a packaged Annex IV dossier on the Enterprise roadmap. Evidence that is cryptographic, not testimonial.
Proof and offline verification are live today; eIDAS and the Annex IV dossier are planned Enterprise modules, not open-build features. Live vs roadmap
Designed for the constraints of banks, insurers, healthcare, and the public sector. We are building IAGA Sentinel Enterprise to turn the open-build mechanisms into dossiers, dashboards, and qualified signatures addressed to EU AI Act, GDPR, and DORA record-keeping obligations, in a separate commercial repository. It is not on sale yet. Leave your email to follow the build and be first in line when early access opens.
Leave your email and we will keep you posted on Enterprise milestones and reach out as early access opens. No sales pitch, unsubscribe anytime.
Generated from the signed chain. Proof is live today; the dossier generator and qualified signatures are planned Enterprise modules, not open-build features.