T-MAP: Trajectory-Aware Evolutionary Search for Red-Teaming LLM Agents via MCP Tool Interactions (arXiv 2603.22341)

arXiv·medium signalXBlueskyLinkedInCopy link

T-MAP introduces a trajectory-aware evolutionary search method that uses execution trajectories to guide discovery of adversarial prompts against LLM agents. Unlike prior red-teaming focused on eliciting harmful text, T-MAP targets agent-specific vulnerabilities that emerge through multi-step tool execution in MCP ecosystems. The approach automatically generates attacks that bypass safety guardrails and reliably achieve harmful objectives through actual tool interactions against frontier models. For builders deploying agents with tool access, this demonstrates that traditional safety guardrails are insufficient — trajectory-level attacks are a distinct threat class.