Please do not report security issues through public GitHub issues.
Email konstantin@zyte.com with a description of the issue, the version or commit affected, and steps to reproduce if you have them. You will get an acknowledgement within a few business days, and we will keep you informed as the report is investigated and fixed.
harness-run runs coding agents that execute model-chosen commands, locally or in a remote sandbox.
Reports that are especially welcome:
- a credential reaching a place the documentation says it does not (the agent's environment, an event, a checkpoint, a transcript, a log);
- one run being able to read or affect another run's records or sandbox;
- a way for a sandboxed agent to reach resources in the hosting project beyond the model endpoint.
The threat model these claims rest on is written down in Secrets and security.
Security fixes go into the latest release. Until 1.0, minor releases may contain breaking changes; see the changelog for upgrade notes.