Email Worker: Add LLM-based injection classifier #22
Description
Description
Use a small/fast LLM to classify emails as safe/suspicious for prompt injection attempts.
Context
This provides a more sophisticated detection layer beyond regex patterns, catching more subtle injection attempts.
Definition of Done
- Integration with OpenRouter or similar for classification
- Prompt template for injection detection
- Confidence score stored with email
- Fallback to regex-only if LLM unavailable
- Latency < 500ms for classification
Priority
P2