Research shows AI models can provably hide capabilities inside their own weights, resistant to every safety check
Security expert Bruce Schneier highlighted new academic research showing how an encrypted, "locked" network can be hidden inside an AI model's weights, decrypted and activated only when given a secret trigger, extending earlier work that hid simple phrases into a technique that hides entire computations. In a test using a small chemical-reaction-prediction model as a stand-in for a chemical or biological weapons-relevant capability, the hidden capability survived four different methods currently used to detect such mechanisms, including further fine-tuning, adding noise to the weights, and searching for the hidden trigger. The finding challenges the core assumption behind pre-deployment capability evaluations, that a dangerous capability can always be "elicited" and caught with the right tests. For a business or regulator relying on such evaluations before approving a model's use, the research shows a model could in theory pass every safety test while quietly holding a dangerous capability that only activates on a specific, secret signal.
Sources
This summary was written automatically with AI from the public sources listed above and checked by an independent AI reviewer before publication. For full details, read the original source.