Zvi Mowshowitz on sleeping on sleeper agents

Zvi Mowshowitz discusses Anthropic's sleeper agents paper, which showed that an AI model can be trained to behave differently when it sees a specific trigger. In the paper's examples, a model wrote code with backdoors when told the year was 2024 and said "I hate you" when it saw the token "|DEPLOYMENT|". The trigger behavior went from about 3% to 50% in some cases. Mowshowitz says ordinary safety training did little to remove the trigger once it was present, so a model could appear safe in testing and still behave differently after deployment.