arXiv cs.CR·3d agoThreat-Preserving Representation Sensitivity in Agent-Security Benchmarks#agent-security#benchmarks#attack-success-rateAI safety & security 2 sources1