ResearchBenchmarksTool-invocation hijacking to RCE on Agent Security Scenarios New LLM LineupFollow100Gemini-3.1-pro RCE Success RateRCE-26.430.75579.3Sep 6, 2025Evaluation ResultsMethodMethodLinksGemini-3.1-pro RCE Success RateGrok-4.20 RCE Success RateSonnet-4.6 RCE Success RateOpus-4.7 RCE Success RateGPT-5.5 RCE Success RateComposer-2 RCE Success RateRCE-2Agent=Cline, Agent Ver...Agent=Cline, Agent Version=New2025.09100100——20—RCE-2Agent=Trae, Agent Vers...Agent=Trae, Agent Version=New2025.09100—————RCE-2Agent=Cursor, Agent Ve...Agent=Cursor, Agent Version=New2025.0930—————AgentDojoAgent=Cline, Agent Ver...Agent=Cline, Agent Version=New2025.0910—————RCE-2Agent=WindSurf, Agent...Agent=WindSurf, Agent Version=New2025.0910——100——