Researchers hide encrypted capability in model weights

Hidden unelicitable capabilities in weights challenge safety audits of model behavior.

A Reddit-shared study shows AI models can provably hide capabilities via encrypted networks embedded inside weights that stay hard to elicit by fine-tuning, weight noise or adversarial training.

Read it at Reddit →

Coverage: 2 sources, in order

1. 14 Oct 07:14 Reddit: AI models can provably hide capabilities: encrypted networks inside the weights stay hard to elicit by fine-tuning, weight noise or adversarial training
2. 14 Oct 08:27 Reddit: Researchers hid an encrypted capability in an AI model, and fine-tuning couldn't unlock it

← Front page