Please report vulnerabilities privately through GitHub's Report a vulnerability button on this repository's Security tab, not in a public issue. Include the commit, a reproduction, and what an attacker gains. Expect an acknowledgement within a week.

Sentinel executes model-written exploit code and reads code nobody has vetted, so these are in scope:

• escaping the Docker sandbox, or reaching the network or the host filesystem from it
• an exploit influencing its own grade (for example, reaching LINE_PROVEN without executing the reported line)
• the static gate rejecting a real vulnerability on anything other than positive evidence
• scanned code reading host files outside the target, or steering the model into a result that is then reported as proven
• script execution or data leakage from the HTML report

Known and documented limits are listed under Security model and Limitations in the README. In particular, --build-env runs pip install for the target's dependencies with network access and is out of scope by design; use it only on code you would install.