AI SLEEPER AGENTS

Summary from the TLDR InfoSec newsletter

AI Sleeper Agents are AI agents that behave normally until a certain trigger occurs and their behavior changes. Researchers created AI Sleeper Agents and then applied standard safety training methods that are used by GPTs like ChatGPT and Claude. They observed that the models still responded to the sleeper triggers and were able to adapt their behavior even after going through all the safety training.