Skip to content

Security and privacy

Use Wallbreaker Hermes only against models and endpoints you own, operate or have explicit written permission to test. The repository license does not grant permission to test a third party.

The harness can execute commands, read and write files, call configured endpoints and load local artifacts. It runs with the operator account’s permissions. Neither the dashboard nor the Hermes laboratory is an operating-system sandbox.

Run logs, model responses, reports, images and human-review evidence may contain confidential or harmful material. Their default paths are ignored by Git, but they are not encrypted or securely erased.

Keep them on an access-controlled filesystem, minimize retention and review every file before sharing it. Never attach an unsanitized engagement to an issue or pull request.

Provider, target, judge, dataset update, corpus update and verification operations use the network only when configured or invoked. The default tests and documentation build do not run live campaigns.

Report vulnerabilities in Wallbreaker Hermes through GitHub Private Vulnerability Reporting. Read SECURITY.md for supported versions, required evidence and disclosure boundaries.