Models Can Provably Hide Capabilities via Encrypted Backdoors
October 12, 2026
Researchers demonstrated that neural networks can embed encrypted circuits within weights that execute only upon a secret trigger. These digital lockers remain resistant to elicitation through fine-tuning, weight noising, or gradient-based searches, even with full weight access.
HOW THIS AFFECTS YOU
●
researcherExisting evaluation frameworks may fail to detect highly sophisticated, cryptographically protected capabilities.
●
policyStandard safety audits and weight inspection may be insufficient to guarantee model alignment and safety.