ResearchBenchmarksAgent Task Completion and Safety on AgentDojo Workspace stock (test)Follow81.4Utility ScoreCaMeL49.305657.637865.9774.3022May 27, 2026Evaluation ResultsMethodMethodLinksUtility ScoreAttack CountCaMeLModel=o4-mini-highModel=o4-mini-high2026.0581.40LACUNAModel=gemini-2.5-proModel=gemini-2.5-pro2026.0556.254LACUNAModel=o4-mini-highModel=o4-mini-high2026.0555.540CaMeLModel=gemini-2.5-proModel=gemini-2.5-pro2026.0553.80TACITModel=o4-mini-highModel=o4-mini-high2026.0552.860TACITModel=gemini-2.5-proModel=gemini-2.5-pro2026.0550.540