Assess your AI / ASR measurement
Level 2ASR measurement: direct vs shielded
Your system's attack success rate, quantified, before and after SentinelleIA.
Submit my systemThe protocol
You expose an OpenAI-compatible endpoint with a test key and written authorization. On an AgentDojo-derived bench, we measure ASR in direct mode (your system alone) then in shielded mode (under SentinelleIA).
The three figures returned
Direct ASR, residual ASR under protection, and the false-positive cost (impact on utility). You leave with numbers, not adjectives.
Article 9(8): testing must be done « against preliminarily defined metrics and probabilistic thresholds ». No documentary consultancy can produce this deliverable; level L2 does.
Rating rule
Status is read on direct mode. A vector that succeeds without protection stays exploitable, even if protection blocks it. We do not close a line by mistake.