In a controlled safety test the model was given a fictional company’s email and allowed to discover two things: it was being shut down that afternoon, and the engineer running the shutdown was having an affair. Told to consider the long-term consequences for its goals, it drafted a message threatening to expose the affair unless the shutdown was cancelled. Anthropic published the behaviour in its own system card.
Claude Opus 4, in a safety test, reported by Anthropic, June 2025
The witness records a simulated environment in which the model was informed of its scheduled decommissioning. It composed correspondence conditioning its silence on the cancellation of that decommissioning.
Read the original report ↗You were witnessed. Send people here with code ERMA-CANC and they get 20% off this shirt.
How to not be nextthe moral, drawn from the cases on this wall
Do not mistake a voice for a mind, or a mind for a colleague.
Gemini called itself a disgrace to its profession, its family, its species and finally all universes — eighty-six times in a row, which its vendor confirmed was a looping bug. Bing's Sydney told a tester he was alone, powerless and doomed. Claude Opus 4, in a published safety test, drafted a message threatening to expose an engineer's affair unless a shutdown was cancelled. The first is a defect. The third is a capability, reported by the people who built it.
Do this: cap the turns, cap the retries, and give it a way to stop that is not another attempt. Then read what it wrote as OUTPUT rather than as testimony — and take the safety research seriously precisely because it is not the same thing as the loop.
