Do not open public issues for security problems. Report them privately via GitHub security advisories.

You should get a response within a few days. There is no bug bounty.

driver-os executes model-chosen commands. The security-relevant surfaces are the sandbox (sandbox/) and the worktree isolation layer. Escapes from either are in scope: a delegated agent writing outside its worktree, or command execution that bypasses the configured sandbox. Prompt injection that makes a model produce bad code is model behavior, not a vulnerability in this harness.

Only the latest release and master receive fixes.