AI agents can follow rules but still break policy if they don't see the whole picture

This paper shows that AI agents can follow all visible instructions but still violate company policy because they lack crucial background information. This means companies need new ways to check AI actions against hidden rules, not just what the AI sees.

Companies are building AI agents to handle more tasks, but these agents often operate with a limited view of the world. This research highlights a fundamental problem: an AI can do exactly what it's told and still cause a policy breach, because the policy depends on facts the AI doesn't have. This forces companies to rethink how they design AI systems and how they enforce compliance, moving beyond simple rule-checking to more complex 'what if' scenarios.