On March 31, 2026, Anthropic, the safety-focused AI company behind Claude, accidentally leaked over 500,000 lines of proprietary source code for its Claude Code tool through a public npm package update. This incident, the second such breach in a year, exposed nearly 2,000 TypeScript files via a misincluded debugging file in version 2.1.88, which linked to a publicly accessible zip archive on Anthropic's Cloudflare storage.Security researcher Chaofan Shou quickly spotted the error, sparking rapid mirroring on GitHub where repositories amassed thousands of stars before takedowns.  The leak revealed Claude Code's full architecture, including 44 feature flags for unreleased capabilities like a "persistent assistant" that runs in the background even when users are inactive. Other hidden gems included session review for performance improvement across conversations, remote control from mobile devices, and a roadmap toward longer autonomous tasks, enhanced memory, and multi-agent collaboration. Developers also uncovered internal tools, prompts, and even a "pet system" codenamed Buddy with species and rarity tiers, hinting at gamified enterprise features.  Anthropic swiftly responded, calling it "human error" in a release packaging issue, not a security breach, with no sensitive data exposed. The company issued over 8,000 DMCA takedown requests to platforms like GitHub, removing thousands of forks within days. Claude Code creator Boris Cherny confirmed a skipped manual deploy step caused the mishap, and Anthropic pledged process improvements to prevent recurrence.  This incident underscores vulnerabilities in AI firms' deployment pipelines, especially for a lab positioning itself as security-conscious amid IPO preparations. Competitors now gain insights into production-grade AI coding agents, potentially accelerating their own developments in agent orchestration and tools. While unlikely to derail Anthropic's $340 billion valuation, it highlights how securing AI systems rivals defending against AI-powered threats.  Ultimately, the Claude Code leak serves as a stark reminder for the AI industry to fortify internal safeguards as innovations race ahead. It boosts hype around Anthropic's capabilities while exposing the human element in high-stakes tech releases. As external developers reverse-engineer remnants, the focus shifts to ethical use and robust verification in open-source ecosystems.

Comments

「30分でClaude Codeを極める」 （Anthropic公式ワークショップ / Boris Cherny登壇）皆さん、こんにちは。私はBorisです。AnthropicのTechnical Staffの一員で、Claudcodeを作成しました。今日はClaudeCodeを使う上での実践的なTipsとTricksについてお話しします。歴史や理論の話はあまりせず、非常に実践的な内容にします。本題に入る前に、簡単に挙手をお願いできますか？ ClaudeCodeをすでに使ったことがある方、挙手をお願いします！（挙手を見て） おお、いい感じですね！ たくさん挙がってて嬉しいです。挙手され

YouTube video by Alex Kantrowitz

体验了下新版的 ZCode+Kimi/DeepSeek ，感觉在国模上比 ClaudCode 好用，最近腹泻式更新，国产也站起来了

VSCode 重度用户，之前 AI 用的不多，用了一段时间 Cursor ，感觉还不错，但不花 20 美元连 BYOK 都用不了，折腾过 Cline ，比 Cursor 差不少。最近 VSCode 更新后 Chat 功能跟 Copilot 已经解耦，甚至不登录 Github 账号也能用，体验上很像 Cursor 了。 它在 AI Coding 工具中处于什么水平？网上搜了一下，似乎都是对比 ClaudCode 和 Codex ，很少拿 VSCode Chat 对比的，但感觉挺好用了呀，用相同模型的话，比前两者差在哪里？

著者：ClaudCode研修室(著)　個人出版　2026/3/29(日)配信

著者：言能 匠(著)　個人出版　2026/3/16(月)配信

Blog personal de Chema Alonso ( https://MyPublicInbox.com/ChemaAlonso ): Ciberseguridad, IA, Innovación, Tecnología, Cómics & Cosas Personasles.