Claude's leaked source code is absolutely UNHINGED!

TL;DR

Anthropic accidentally leaked 512,000 lines of Claude Code source code twice, revealing hidden features like undercover mode, autonomous agents, and a poison pill for competitors.

Key Points

• 1.Anthropic leaked its entire Claude Code source code twice via the same mistake. On March 31st, 2026, a missing single line in a settings file exposed 512,000 lines of code via an unprotected zip file — the identical error had already occurred 13 months earlier on Claude Code's launch day in February 2025.
• 2.Undercover mode is the most controversial discovery. 90 lines of code instruct Claude to hide all AI authorship — stripping Anthropic fingerprints, model references, and AI-sounding commit messages when employees contribute to public open-source projects, with 'no force off' option.
• 3.Hidden unreleased features reveal an autonomous AI workforce roadmap. 'Chyros' is an always-on background agent filing bugs and pushing updates while you sleep; 'Ultra Plan' offloads project planning to a remote Claude for 30 minutes; 'Tungsten Tool' gives Anthropic employees-only full virtual terminal control.
• 4.A poison pill actively injects fake decoy data into every request. To deter competitors from reverse-engineering Claude Code's behavior by intercepting traffic, fake features are mixed with real ones on every API call — undisclosed to users.
• 5.The codebase quality raises serious concerns about AI-written software. 100% of the head of Claude Code's contributions were written by Claude itself; key files have 5,000+ lines with 219 imports, forming 'god objects' optimized for AI efficiency rather than human maintainability.
• 6.Security researchers found a prompt injection vulnerability in the conversation summarizer. A hidden second Claude silently compresses long conversations, but Straker security firm found malicious instructions planted in project files survive the summary and get laundered into trusted context — attackers need only plant a file.