SentinelMesh is my research project for securing AI agents in production: a mesh layer that instruments, evaluates, and responds to risky behavior in real time.
The problem
Agents call tools, read files, and hit APIs with minimal human oversight. Traditional WAFs and endpoint tools weren't designed for:
- Prompt-level attacks that change intent mid-session
- Tool chains that escalate privilege step by step
- Opaque model outputs that leak context
The approach
SentinelMesh sits alongside your agent runtime:
Agent → Mesh (policy + detect) → Tools / APIs
↓
Audit & alerts
Three pillars:
- Instrument — structured events for every tool call and policy check
- Evaluate — rules plus detectors for injection, exfiltration, abuse
- Respond — block, alert, or quarantine with full audit trail
Status
Active research. See the projects page for architecture notes and links.