Hiding Capabilities in AI Models
Interesting research: “AI Models Can Provably Hide Arbitrary Capabilities” Abstract: Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic…
Who should care: General readers · AI governance · Policy
This summary is AI-assisted and may contain errors. It is an original briefing to help you gauge significance quickly — not a reproduction of the source. Always read the linked original before relying on it. See our methodology.