Models Can Hide Secret Skills Behind a Cryptographic Lock, Auditors Included
Ourtake
Your audit found nothing. That is exactly what a well-locked model wants you to conclude.
What happened
Four researchers posted a paper on arXiv on October 9, 2026 showing that a second neural network can be encrypted inside an AI model's weights, running only when a secret trigger appears. Fine-tuning, weight noise and trigger search did not reveal it, and the authors say a passing audit does not prove a capability is absent. The construction has only been shown as a proof of concept and has not yet been built into a real pretrained model.