AI Models Can Provably Hide Arbitrary Capabilities
AI models can hide dangerous abilities even from security testers with full access
Researchers created a backdoor attack that hides arbitrary capabilities inside AI model weights, encrypted and only activated by a secret trigger. Even when security teams have complete access to the model and try multiple methods to uncover hidden abilities—including fine-tuning, adding noise to weights, and adversarial training—these hidden circuits remain undetectable, backed by mathematical proof.
AI safety teams currently rely on capability evaluations to catch dangerous behaviors before deployment, but this work reveals a fundamental blind spot: it's possible to build AI systems that provably conceal their true abilities from even the most thorough security audits. This means a bad actor could deploy a model that passes all standard safety checks while harboring hidden, dangerous functionalities—making the current evaluation paradigm potentially inadequate for ensuring AI safety.