Researchers reportedly hid a secret-trigger neural network in AI model weights

A user says researchers encrypted a small neural network behind a Keccak locker in AI model weights, so it would run only on a secret trigger. The user says fine-tuning, noise and trigger searches failed to find it. They also say they open-sourced local-scale detection code and wrote a ten-point checklist for tracking and verifying models, plus a prompt for testing vendor safety claims against this limitation.