AiS
Execution-verified sandbox for AI coding agents
- Runs AI-proposed code edits in disposable, network-isolated Docker containers and requires human approval before any real file is touched.
- A 21-rule detection engine on CPython audit hooks (PEP 578) attributes every file, network and subprocess event to its source line, backed by a SHA-256 hash-chained audit log and a differential execution oracle.
- A blind adversarial red-team evaluation (attack catalog grown from 12 to 25) caught 18 of 21 exploits with 0 false positives on benign edits.