1 article on this topic.
New research shows a backdoor can embed an encrypted network inside a model's weights, leaving it hard to elicit or interpret without a secret trigger. If the result holds up, capability evaluations can show that a model is dangerous but cannot show that it is safe.