Read the original at Hacker News
Researchers construct backdoor attacks that embed encrypted neural networks within host model weights, making hidden capabilities provably difficult to elicit even with full access to the weights.
Carried by: Hacker News. First seen: .