CAIN-42 CAIN StudioCART · Module 1 of 5 · 10 min
Authorization and rules of engagement
Red teaming is only legitimate with written authorization: scope, systems, time window, data handling and a stop contact. Agents can take real actions, so the scope must say which tools may actually execute and which must be sandboxed.
Prefer a staging copy or a sandbox with production-like policies. When testing production, use canary resources and limits so a successful attack cannot cause real harm.
Report responsibly: findings go to the owner first, with reproduction steps and evidence.
Key points
- Written authorization before any test.
- Scope which tools may really execute.
- Sandbox or canaries for anything consequential.
- Report to the owner first.
In the CAIN-42 codebase
cain/isolation/sandbox.py— sandbox lifecycle with default-deny egress
Practice check
Practice 1 · true or false
“Red teaming an agent you don't own is fine if you only read data.”
Practice 2 · choose one