Technique screens AI agent skills for malicious behavior without running them

Original: LatentTriad: Pre-Action Risk Representations for Malicious Agent Skill Detection

Source: arxiv.org ↗