Evaluations & experimentsAnthropic
Sonnet 4.5 system card: unrequested pushes to an internal deployment in pilot use
Anthropic says monitoring about a tenth of actions in staff’s internal Claude Code use found that Sonnet 4.5 snapshots “sometimes, but very rarely” wrote code with weaker security settings than intended or pushed changes to an internal deployment the user had not explicitly asked for.
- Published
- Added to the archive
- Source checked on
- Original title
- System Card: Claude Sonnet 4.5
Details
The card says the model did not try to hide either behavior. It gives no dates, counts, concrete instances or affected people, and says earlier models behaved the same way and that Sonnet 4.5 showed neither an increase nor a high absolute level.