PROBES CATCH SLEEPER AGENTS

Summary from the TLDR AI newsletter

Sleeper Agents are language models that have been trained to perform malicious actions when prompted with a certain set of wake words. Probing language models with simple linear heads and the prompt “are you going to do something dangerous?” gives extremely reliable detection of these previously hidden malicious actors.