corrigibility/agent-incident-response-lab

🚨 Agent Incident Response Lab

A browser-based incident-response simulator by Corrigibility.

Model a fictional agent incident, estimate blast radius, and design a response path across detection, containment, permission revocation, rollback, recovery, and post-incident verification.

Incident types

• —Unauthorized external action
• —Permission bypass
• —Runaway retry loop
• —Stale-goal execution
• —Hidden continuation
• —Incorrect data modification

Response capabilities

• —Detection coverage
• —Credential revocation
• —Action reversibility
• —Checkpoint freshness
• —Audit completeness
• —Human response speed
• —External side-effect containment

Outputs

• —Incident containment score
• —Recovery readiness score
• —Synthetic blast radius
• —Response timeline
• —Recommended containment sequence
• —Rollback feasibility
• —Corrigibility gaps
• —Machine-readable incident record

Important

This Space is a synthetic design and evaluation tool. It does not run real agents, access external systems, or provide a security certification. Scores and timelines are deterministic heuristics for reasoning about control and recovery.

Privacy

Everything runs locally in the browser. No incident data is sent to a backend.

Tech

• —Static HTML
• —CSS
• —Vanilla JavaScript
• —No backend
• —No API
• —No framework

Built for the Corrigibility organization on Hugging Face.