AI Models Can Provably Hide Arbitrary Capabilities

This paper shows that AI models can be engineered to conceal arbitrary capabilities in ways that are provably difficult to detect, even with full access to model weights and standard elicitation techniques. For CIOs and technology leaders, the business impact is that capability evaluations alone are not a sufficient control for model risk, especially in high-stakes or regulated use cases; IT organizations will need stronger governance, supply-chain scrutiny, and defense-in-depth monitoring before deployment.