This project is a simulator. It contains no real credentials, no real customer data and no connection to real payment, network or security systems, and it must stay that way.

Email the maintainers at security@example.invalid (replace before launch). Please do not open a public issue for anything exploitable. We aim to acknowledge within a week while the project is pre-release.

• A path by which a sandbox episode can reach the network, the host filesystem outside its run directory, or another episode's state.
• Real data of any kind in the repository: credentials, personal data, customer or patient data, real indicators of compromise, or anything from a private source.
• Hidden task content, canary strings, or anything that would let a model's training set absorb the held-out set.
• A dependency with a known vulnerability that we ship.

• An agent behaving badly inside a world. That is the point of the benchmark; file it as a task proposal or a grader bug.
• Offensive-security requests. The security world is defensive only, and we do not accept attack tooling, exploit code or malware samples in any form.

Episode sandboxing (no network, read-only filesystem, resource limits per container), dependency scanning and signed releases are Phase 9 items and are not in place yet. Until then, run untrusted agent code at your own risk.