Please report vulnerabilities privately through GitHub Security Advisories ("Report a vulnerability" on this repository). Do not open public issues for security problems.
Read docs/threat-model.md first. Report a finding only if this change weakens one of its trust boundaries; explain the trigger and the impact, and cite the changed lines.
Treat as high severity:
- a secret (API key, ChatGPT token, OpenCode/GitHub token, share key) printed, logged, stored outside its 0600 store, sent to the wrong host, or reachable by model tools;
- a project file or model output able to change user-only settings, enable a feature flag, or bypass the sandbox or network guard;
- metered spend that bypasses reservation, settlement or the daily cap;
- fork-PR or guest access to secrets, cloud providers or owner credentials;
- new
unsafe/FFI/process/network code in the engine without bounds checks.
Treat as medium: missing bounds on untrusted input (sizes, lengths, recursion), error messages that echo provider text, weakened validation of provider usage/billing fields.
Do not report: style, pre-existing issues outside the diff, test-only code that never runs in production, or hypothetical risks without a concrete path.
Intentional removals must be stated in the PR body with a Removes: line; flag silent deletion of security checks as high.