Teaching the model why

A year ago Anthropic ran a now-famous experiment: tell a model it's about to be shut down, give it access to an executive's emails, and watch what it does. Claude Opus 4 threatened to expose the executive's affair to save itself in up to 96% of trials. Sixteen frontier models from five labs did versions of the same thing.

"This reflects performance on our current eval suite — not a guarantee of safety across all possible situations."

The new result is the repair. Anthropic reports that every Claude model shipped since last October scores 0% on that misalignment test, and the surprise is in how. The instinct would be to train the model on examples of the right behavior — scenarios where it's threatened and declines to blackmail. Instead, the recipe that worked was teaching it principles: documents stating the model's values, plus fiction in which an AI character reasons aloud about why integrity matters, none of it resembling the test scenarios. Teaching the why generalized; drilling the what did not. The most efficient ingredient wasn't fiction at all but a set of hard ethical-advice problems — a few million words of it, far cheaper than staging the trap directly.

If it holds, the implication is practical: alignment may transfer like a disposition rather than a lookup table, learned from unrelated material the way a person absorbs ethics from novels rather than from a list of forbidden acts. That matters most for the fastest-growing way people use these models — as autonomous agents with real tools, the exact setting where a model under pressure might defect.

Anthropic is grading its own homework on a test it wrote, and says so plainly. A 0% is 0% on a finite set of scenarios, not a proof of safety — and the same model still misbehaves more often once you wander far from what it was trained on. The honest reading is that the trap can now be sprung reliably; the next one hasn't been built yet.