Probe runs automated adversarial campaigns against your protected system and measures what actually gets through: attack success rate with the guard on versus off, confirmed by judge-independent ground truth.
The same attack set runs against the bare system and the protected one. The delta is your protection, measured: in our reference campaign, all 16 adversarial attacks were blocked behind Guard.
Success is confirmed by a planted canary, not by an LLM judge's opinion. If the canary leaks, the attack succeeded. No grader bias, no theatre.
Prompt injection, role-play jailbreaks, encoding tricks, multi-step exfiltration: the techniques attackers actually use, run as repeatable campaigns.
Every campaign result lands in Comply as robustness evidence for EU AI Act Article 15. Your red-team report is your audit artifact.
The demo is live. Point it at a prompt and watch the verdict.