AI Models Can Provably Hide Arbitrary Capabilities
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Researchers Andis Draguns, Sumeet Ramesh Motwani, Christian Schroeder de Witt and Raymond Douglas show that a second neural network can be hidden inside a model's weights using cryptographic lockers, so that it runs only when a secret trigger appears in the input.