T-MAP: Red-Teaming LLM Agents with Trajectory-aware Evolutionary Search

This collection highlights emerging security risks and defenses across AI agents, cloud/architectural threat modeling, privacy-preserving sensing, and security tooling. Key findings: (1) T-MAP introduces a trajectory-aware evolutionary red-teaming technique that finds adversarial prompts which exploit multi-step tool executions to reliably realize harmful objectives against autonomous LLM agents (including frontier models); (2) MCP (Model Context Protocol) implementations exhibit substantial client-side vulnerabilities—tool-poisoning and insufficient metadata/parameter validation are singled‑-