AI Models Can Provably Hide Arbitrary Capabilities
AI summary
Editorial policyResearch shows that the assumption of invoking hidden functions collapses under adversarial conditions, demonstrating a method to embed encrypted neural networks via backdoor attacks.
Key points
- •It showed that capability evaluations rely on the assumption that hidden functions can be invoked, which does not hold under adversarial conditions.
- •It constructed a backdoor attack that decrypts and executes only upon receiving a secret trigger.
- •By using digital lockers, it made it provably difficult for hidden circuits to be invoked or interpreted.