flâneur

Simple probes can catch sleeper agents \ Anthropic

anthropic.com · 3,523 words · saved by 1 readers

This “Alignment Note” presents some early-stage research from the Anthropic Alignment Science team following up on our recent “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training” paper. It should be treated as a work-in-progress update, and is intended for a more technical audience than our typical blog post. This research makes use of some simple interpretability techniques, and we expect to share more results from collaborations between our Alignment and Interpretability teams soon. In this post we present “defection probes”: linear classifiers that use residual stream activations to predict when a sleeper agent trojan model will choose to “defect” and behave in accordance with a dangerous hidden goal. Using the models we trained in “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training”, we show that linear detectors with AUROC scores above 99% can be created using generic contrast pairs that don’t depend on any information about the

Alignment Interpretability Simple probes can catch sleeper agents Apr 23, 2024 This “Alignment Note” presents some early-stage research from the Anthropic Alignment Science team following up on our recent “ Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training ” paper. It should be treated as a work-in-progress update, and is intended for a more technical audience than our typical blog post. This research makes use of some simple interpretability techniques, and we expect to share more results from collaborations between our Alignment and Interpretability teams soon. Sum

related reading