SeClaw: Spec-Driven Security Task Synthesis for Evaluating Autonomous Agents

Researchers introduced SeClaw, a framework for generating security tasks from structured risk specifications.

SeClaw evaluates autonomous LLM agents in a standardized Docker-based testbed for reproducible safety scenarios.

The framework broadens coverage across resource, task, environment, and agent-behavior risks.

It judges unsafe trajectories, not just final outputs, making safety assessment more systematic and scalable.