Researchers propose peeking inside LLM 'black box' for safety signals

Researchers suggest identifying cognitive elements in LLMs that signal potential unwanted actions, aiming to improve AI safety by interpreting internal model states.