This simple approach works here because prompts that induce dangerous behavior are salient in the internal state of these sleeper agent models. This is likely due to the way they were fine-tuned. How effective the probing techniques will be in practice remains an open question.

→ View original post on X — @anthropicai