The sandbox and worker tiers need a durable threat catalog, not one-off escape tests buried in happy-path files.
Problem
Infinite loop, process spawn, filesystem read, and env exfil attempts should be named, expected, and regression-tested on both tiers. Residual constructor-walk risks should stay visible instead of implied fixed.
Intended fix
- Add
src/threats.ts catalog with per-tier expected outcomes
- Run containment tests for loop / fork bomb / fs / env
- Pin residual constructor-walk behavior
- Write
docs/security.md for the trust model
The sandbox and worker tiers need a durable threat catalog, not one-off escape tests buried in happy-path files.
Problem
Infinite loop, process spawn, filesystem read, and env exfil attempts should be named, expected, and regression-tested on both tiers. Residual constructor-walk risks should stay visible instead of implied fixed.
Intended fix
src/threats.tscatalog with per-tier expected outcomesdocs/security.mdfor the trust model