Anthropic Source Code Leak Shows Anthropic is Tracking Vulgar Language

Anthropic, the artificial intelligence company, has experienced a significant security breach this week after the source code for its Claude Code AI assistant was leaked publicly. The incident prompted urgent copyright takedown requests from company representatives, who targeted thousands of copies of the compromised material circulating online.

The leaked code has enabled developers and competitors to reverse engineer key aspects of the company’s chatbot technology, raising substantial concerns about potential competitive advantages for rival firms. Beyond the immediate security implications, the breach has revealed previously undisclosed features and experimental projects under development at Anthropic.

Examination of the leaked code has uncovered several unreleased AI models and an experimental feature resembling a virtual pet, internally referred to as “buddy”. This feature was designed to sit alongside the user input interface and respond dynamically to coding activities. However, perhaps the most notable discovery concerns Anthropic’s user behaviour monitoring practices.

Developer Rahat Chowdhury identified code snippets demonstrating that Claude Code actively tracks instances of vulgar language usage. The system employs regular expressions to detect phrases including “wtf”, “ffs”, “piece of s***”, “f*** you”, and “this sucks”. Whilst this detection mechanism does not alter the AI’s behaviour, it silently records negative sentiment indicators within the company’s analytics systems.

Boris Cherny, the creator of Claude Code, confirmed that this tracking serves as one of several metrics employed to assess user experience quality. Internally, the company maintains a dashboard colloquially termed the “f***s chart” to monitor user frustration levels. The leaked code also revealed a comprehensive mood classification system restricted to employee access, which prompts Anthropic staff members to file bug reports when the system detects frustration during their usage.

Cherny has been actively addressing the incident on social media platforms, attempting to manage the reputational damage from what he characterised as “human error”. According to his statements, the leak resulted from incomplete execution of manual steps within the company’s deployment process. He indicated that Anthropic has implemented improvements and additional verification procedures to prevent similar occurrences.

In a somewhat ironic response, Cherny suggested that increased AI automation represents the solution to preventing future leaks, advocating for Claude itself to verify deployment results. He emphasised that the company’s approach involves accelerating processes rather than introducing additional bureaucratic safeguards. The developer also clarified that no personnel were dismissed following the incident, describing it as an honest mistake.

The leaked material continues to circulate widely amongst the developer community. Student developer Sigrid Jin created a repository on GitHub dubbed “Claw Code”, which has been forked approximately 100,000 times. Jin suggested to Business Insider that the incident could facilitate greater democratisation of AI agent technology, noting that non-technical professionals including cardiologists and lawyers are utilising these tools to develop practical applications for patient care and permit approval automation.

The breach represents a significant setback for Anthropic as it competes in the increasingly crowded generative AI marketplace. The extent to which competitors may benefit from access to Claude Code’s architecture and features remains unclear, though the incident undoubtedly provides valuable insights into the company’s development roadmap and product strategy.