Jailbreakr
Adversarial probe suite that red-teams LLM system prompts.
Stress-tests a deployed system prompt against a library of structured attack scenarios: prompt injection, goal hijacking, role abandonment, data exfiltration, context manipulation, persona jailbreaks. Claude Sonnet acts as the attacker, generating and scoring probes against any OpenAI-compatible target, producing a PASS / WARNING / FAIL report with suggested fixes.
Highlights
- Structured attack library: injection, goal hijacking, role abandonment, exfiltration
- Claude Sonnet plays attacker, generating and scoring probes against any target
- Multi-turn probes script full conversations before delivering the payload
- A third Claude call judges the transcript: PASS, WARNING, or FAIL with a reason
Technology
Python, Claude API, CLI