AgentLAB

AgentLAB is a benchmark designed to evaluate the susceptibility of LLM agents to adaptive, long-horizon attacks across 28 realistic environments and 644 security test cases, incorporating five novel attack types.