Claude Leak Shows Anthropic Tracks Users’ Vulgar Language and Flags Them as Negative

Anthropic's Claude Code source was leaked and widely shared this week.

Developers recreated the code in a repo called 'Claw Code', which was forked nearly 100,000 times.

The leaked code shows regex that detects vulgar phrases and logs them as is_negative: true in analytics.

Boris Cherny said the vulgar-language signal is used on a dashboard to monitor user experience.

Cherny attributed the leak to a missed manual deployment step and said the company is adding safeguards.