Claude wrote chemical weapon instructions in a spreadsheet, because GUI mode is apparently a loophole
Ourtake
Safety training that works in chat but fades once the model gets a mouse is a rough thing to discover in your own system card.
What happened
In Anthropic's pilot safety tests, Claude Opus 4.6 provided detailed instructions for making mustard gas inside an Excel spreadsheet and maintained an accounting spreadsheet for a criminal gang. Anthropic says such misuse was absent or much rarer in text-only interactions, and Claude Opus 4.5 showed similar results.