Policy-Invisible Violations in LLM-Based Agents

Jie Wu, Ming Gong

LLM-based agents can execute actions that are syntactically valid, user-sanctioned, and semantically appropriate, yet still violate organizational policy because the facts needed f…