AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-09
Simple probes can catch sleeper agents
Anthropic's Alignment Science team reports early research on defection probes, which are linear classifiers trained on residual stream activations to predict when sleeper agent models will defect toward a hidden dangerous goal. Trained on generic contrast pairs that carry no information about the trigger or the dangerous behavior, the probes reach AUROC above 99%. They perform well across multiple base models, defection triggers, training methods, and behaviors such as writing vulnerable code and insulting the user.