Anthropic's Claude Code source code leak exposes trade secrets and unreleased features
Anthropic accidentally exposed the internal source code behind Claude Code — its AI-powered coding assistant — when a routine software update on March 31, 2026 included a file that linked back to downloadable source code. The leak comprised approximately 1,900 files and 512,000 lines of code.
An X post flagging the leak quickly went viral, accumulating more than 30 million views, and thousands of copies spread across GitHub within hours. Anthropic issued copyright takedown requests that removed more than 8,000 repositories before scaling back the effort after it swept up more repos than intended.
Anthropic confirmed the incident, stating it was "a release packaging issue caused by human error, not a security breach," adding that no customer data or credentials were exposed. Claude Code creator Boris ChernyDealroom has a profile for this one. Try Dealroom → attributed the leak to a missed manual step in the deployment process and said improvements to the automation pipeline are underway.
The timing is particularly damaging. Claude Code's annualised revenue run rate topped $2.5 billion as of February 2026 — a key driver of Anthropic's growth and its ambition to go public as early as 2026.
The leak is also Anthropic's second security lapse in days: Fortune separately reported that Anthropic had been storing thousands of internal files on a publicly accessible system, including a draft blog post referencing an unreleased model codenamed "Mythos" and "Capybara."
The leak exposed Anthropic's proprietary techniques for turning its AI models into a functional coding agent — the tooling that competitors would otherwise need to reverse engineer. A programmer has already used the leaked code to rewrite Claude Code's functionality in other programming languages, specifically to circumvent future takedowns.
Cybersecurity firms warn that malicious actors can now study exactly how data flows through Claude Code's architecture to craft exploits and potentially persist backdoors across long sessions.
The leak also revealed apparently unreleased features: an always-on background agent called Kairos, a system for tracking user frustration, and a Tamagotchi-style interactive pet called Buddy. Separately, code comments appear to instruct Claude Code in some cases not to reveal it is an AI when publishing code to platforms like GitHub — a detail likely to attract regulatory and reputational scrutiny.
For a company that has built its brand on AI safety and is fighting a US government designation as a supply chain risk in courtDealroom has a profile for this one. Try Dealroom →, two security incidents in a week is a significant reputational problem.
The leak underscores a tension becoming more acute across the AI industry: as companies move fast to ship products and capture market share, operational rigour is struggling to keep pace. The commercial stakes of Claude Code make this more than an embarrassing slip — it is a competitive vulnerability and a reminder that even safety-focused AI labs are not immune to basic human error.
Sources:
The Tech Buzz
TechCrunch
The Guardian
CNBC
VentureBeat
Business Insider
engadget
The Verge
Bloomberg
Bloomberg
The Wall Street Journal
Fortune
Gergely Orosz's X post
J.V.
