[ 03 ] Defense in depth for AI agents
MCP Agent Gateway
A four-layer security gateway that puts strict action boundaries, an audit trail, and human-in-the-loop escalation around agent tool calls.
Role
AI security engineering
Stack
- MCP
- AI security
- Audit logging
- Human-in-the-loop
- Testing
What it delivers
- Combines manifest-poisoning detection, rug-pull diffing, kill-chain detection, and canary tripwires.
- Measures per-layer recall against a 20+ case adversarial test suite.
- Adds tamper-evident audit logging and human-approval escalation for high-impact calls.