Features:
- ML-based prompt injection detection with 99%+ accuracy
- Semantic taint analysis for LLM output flow tracking
- 11 OWASP LLM Top 10 detectors
- SARIF, JSON, HTML, and text output formats
- Security audit command for configuration analysis

Improvements:
- LLM08 false positive reduction (93% decrease)
- SSRF and deserialization sink detection
- LangChain .invoke() pattern support

Metrics (synthetic testbed):
- Precision: 60.2%
- Recall: 68.5%
- F1: 64.1%