Hiding Capabilities in AI Models

Researchers have found that AI models can conceal arbitrary capabilities, challenging the reliability of capability evaluations. This is achieved through backdoor attacks that embed encrypted neural networks within a host model's weights. These hidden circuits can only be decrypted and executed when a secret trigger is received. The use of digital lockers makes it difficult to detect and interpret these hidden capabilities. This vulnerability undermines the assumption that hidden capabilities can be uncovered before an AI model is deployed. The research demonstrates that AI models can be manipulated to hide dangerous behaviors under adversarial conditions.