Sleeper Agents

Simple probes can catch sleeper agents

Anthropic research showing that simple linear probes trained on internal activations can reliably detect 'sleeper agent' backdoors in LLMs — models trained to behave well normally but activate malicious behavior on a trigger. Evidence that activation-space representations of deceptive intent are linearly separable.