Evidence · DiggingBeagle record
HarnessSecurity-Bench: Do Security Mechanisms Really Protect Coding Agent Harnesses?
Oct 6 original study comparing agent harness security settings using 2,500 controlled trials.
- Source role
- primary disclosure
Evidence record
Oct 6 original study comparing agent harness security settings using 2,500 controlled trials.
Claim-level citations (2)
- supportsAuto-approve sharply increased attack effects in a coding-agent harness benchmark: Under the GLM-5.2 baseline, auto-approval increased attack completion from 29.2% to 95.6%.
Abstract, security setting comparison
- supportsAuto-approve sharply increased attack effects in a coding-agent harness benchmark: The study evaluated nine security mechanisms across six coding-agent harnesses with 2,500 runs and deterministic oracles.
Abstract, experiment design and evaluation