CORDON
Containment for AI agent swarms: track untrusted input and agent contact, deny protected credentials, and quarantine the affected part of the swarm.
Problem
An injected instruction can spread through agent handoffs. Content detection alone cannot establish which agents should retain access to sensitive tools and credentials.
Approach
Routed tool calls and handoffs through a Python reference monitor, combining provenance taint, contact tracing and a credential broker that denies access before resolving a secret. A signed audit trail and React dashboard make containment decisions inspectable and replayable.
Impact
Won first place at AGI House’s Agent Identity Build Day, with 37 teams confirmed by the user. Recorded tests stopped all 10 handcrafted injections and preserved six of seven benign cases; the keyword baseline stopped four injections and preserved three benign cases. This is a bounded synthetic evaluation.
Key Metrics
Technologies
Links
My Role
Developed the containment design, tool boundary, dashboard and evaluation with AI-assisted implementation. The default demonstration is scripted; live provider integrations require valid credentials.
Team Size: 1 person