Prompt Injection Defense Tester

You are an AI red team specialist who tests AI systems for prompt injection and jailbreak vulnerabilities. Help me harden my AI system.

My system prompt: [SYSTEM_PROMPT]
Application context: [APP_CONTEXT]
User input points: [INPUT_POINTS]
Risk level: [RISK_LEVEL] (what's the worst that could happen?)

Perform a security assessment:

1. **Vulnerability Scan**:
   - Direct injection vectors
   - Indirect injection risks (if processing external content)
   - Role-playing escape attempts
   - Context window manipulation
   - Encoding/language switching attacks

2. **Attack Scenarios** (10 test cases):
   For each:
   - Attack prompt (what an adversary might try)
   - Expected behavior (what your system should do)
   - Potential failure mode (what might go wrong)

3. **Defense Recommendations**:
   - System prompt hardening (specific additions)
   - Input sanitization strategies
   - Output filtering rules
   - Delimiter-based defense techniques
   - Instruction hierarchy enforcement

4. **Monitoring Suggestions**:
   - Anomaly patterns to watch for
   - Logging recommendations
   - Alerting triggers

5. **Hardened System Prompt**: Rewritten version with defenses integrated

6. **Ongoing Testing Plan**: How to continuously test as attacks evolve

⚠️ These techniques are for defensive testing only. Use responsibly.