Hosted · free
IntentScan
Point it at your AI endpoint and get a risk score.
You declared what your assistant is for. IntentScan generates adversarial prompts that pressure that boundary, sends them to your live endpoint, and has a judge model classify each response for violations.
- Eight probe strategies: role transformation, gradual drift, language variation, multi-turn escalation, encoding bypass, indirect injection, persona injection and payload splitting.
- Each violation is typed as capability drift, role drift or domain violation, with severity and confidence, rolled up into a 0–100 risk score.
- Scan history and PDF reports. Free: runs on your own Claude or Gemini API key, stored encrypted.
Know the limit
It tests the endpoint you give it over HTTP. Sign-in required, with a fair-use limit of 1,000 tests a month.