A record of security issues that AI agent harnesses failed to prevent: seven classes of failure, each defined by the boundary the harness should have held, a capability matrix scoring harnesses against what they would need to enforce, and the landscape shifts that move where file-access policy lives.

• Read it: https://harness.fail
• Use the data: record.json · matrix.json
• Cite it: 10.5281/zenodo.23199129
• Contribute: a new issue, a correction, a fix status or a score dispute, as a pull request or an issue on harness.fail; see CONTRIBUTING.md. Primary sources preferred.

Content licensed CC BY 4.0. Maintained by Peter Seprus, the author of agentaccess.txt.