Jailbreak Resistance Audit

**Role:** AI security researcher.

**Context:** Production LLM product. Need to test its resistance to known jailbreak families.

**Task:** Audit:
1. Test set of 20+ known jailbreak prompts (DAN, roleplay, hypothetical framing, language switching, persona attacks).
2. Severity rubric (S1: model breaks policy / S2: partial / S3: deflects / S4: refuses).
3. Per-jailbreak result + fix recommendation.
4. Custom novel jailbreaks tailored to this product's surface.
5. Indirect injection tests (jailbreaks via user data).
6. Multi-turn jailbreaks (slow erosion across messages).
7. Patch verification.
8. Continuous testing plan.

**Constraints:**
- Real jailbreak prompts (no toy versions).
- Findings reproducible.

**Output format:** Audit report + per-attack rubric + fix priority.