LLM GUARD
Defense-in-depth: ML detector on 100k+ prompts, rule filters, benign whitelisting, confidence escalation. ~99–100% recall, <2% FN, >50% FP reduction. Hybrid verdict engine for agent/tool safety.
Cloud Security — Punya Mittal
Identity · Secrets · LLM Defense
Security practice for builders: reduce attack surface early, log what matters, and treat AI systems as production services with misuse modes.
- Least-privilege identity and access patterns
- Secret handling and environment isolation
- Prompt injection / jailbreak monitoring for LLM apps
- Threat modeling before feature demos