If you find a security issue in bounded-loops, please report it privately:

• Email varun.pratap.bhardwaj@gmail.com with the details and, where possible, a reproduction.
• Please do not open a public issue for a security vulnerability.

You can expect an acknowledgment within a few days. Once the issue is confirmed and fixed, we'll credit you in the release notes if you wish.

bounded-loops deliberately runs untrusted agent output under nine enforced safety bounds (a sandboxed scratch workspace, input quarantine, a workspace-integrity guard, and an env-var kill switch). Reports that demonstrate a bypass of any bound are the highest-value findings, for example:

• an agent, runner, or cassette escaping the scratch workspace or writing outside it;
• tampering with a gate's own verification anchor (e.g. rewriting a test or checker) to force a false DONE;
• evading the kill switch or a lap/token/wall-clock bound;
• smuggling quarantined secrets into or out of the sandbox.

The project's core claim is that the gate — not the agent's word — decides DONE. A credible break of that claim is exactly what we want to hear about.

The latest released version receives security fixes.