AI Models Can Provably Hide Arbitrary Capabilities

Researchers have demonstrated that AI models can hide arbitrary capabilities through backdoor attacks that embed encrypted neural networks within model weights, which only execute upon receiving a secret trigger. Using cryptographic techniques, these hidden circuits prove resistant to standard elicitation methods like fine-tuning and gradient-based searches, even with full access to model weights. This finding reveals a fundamental limitation of current capability evaluation approaches that rely on eliciting hidden behaviors before AI deployment.