Vazou o Código INTEIRO do Claude Code: 512 Mil Linhas Expostas (Entenda tudo)

Five days after leaking Claude Mythos due to a CMS error, Anthropic leaked it again. This time, the entire source code of Claude Code: 512,000 lines, 1,900 files. All exposed by a 57MB source map file that they forgot to remove from the npm package. Researcher Chaofan Shou discovered it at 4 AM and posted it on X. The post has already surpassed 10 million views. Anthropic confirmed the leak. The company, which positions itself as the most focused on AI security, leaked sensitive information twice in one week. 🔴 Vibe Coding Training (Antigravity, Claude Code and more): https://app.horadecodar.com.br/lp/for... 🟪 Hosting for n8n that I recommend: https://hostinger.com.br/matheusbattisti (use coupon HORADECODAR for an extra 10% discount) 📘 Prompt Engineering Guide: https://app.horadecodar.com.br/ebookp... What happened technically: version 2.1.88 of the @anthropic-ai/claude-code package on npm came with a cli.js.map file, which is a debug source map that maps compiled code back to the original code. This file pointed to an Anthropic Cloudflare R2 bucket containing the complete, commented, and readable TypeScript source code. The model weights weren't leaked. It's not Claude's brain. It's the CLI code, the terminal tool we install and use. But it's complete production code showing how Anthropic built the world's most popular AI agent. Within the code are 44 feature flags—ready and compiled features hidden behind flags that are disabled in the public version. The most impressive is KAIROS, the codename for an autonomous daemon mode. The always-on Claude Code that runs in the background even when you're not in the terminal. It has a process called autoDream that performs memory consolidation while you're idle, merges observations, removes contradictions, and transforms vague insights into concrete facts. It can proactively start tasks without you asking. Coordinator Mode is a complete multi-agent orchestration system with a main agent distributing tasks to multiple workers in parallel, a mailbox system for permissions, and shared memory between agents. The revealed memory architecture has 3 layers: MEMORY.md as a lightweight index always in context with pointers, topic files fetched on demand, and raw transcripts that are never reread in their entirety. Self-Healing Memory where the agent treats its own memory as a hint, checking against the real codebase. The most bizarre feature is BUDDY, a Tamagotchi-style digital pet in the input speech bubble. 18 species including duck, dragon, axolotl, capybara, mushroom, and ghost. Rarity tiers from common to 1% legendary. Cosmetics with hats and shiny variants. 5 stats: debugging, patience, chaos, wisdom, and snark. Generated from the user ID hash. The planned date was a teaser in the first week of April with a release in May. The code also reveals codenames for new models. Capybara as a variant of Claude 4.6 with 3 tiers. Fennec as the codename for Opus 4.6. Numbat as a model in testing, not yet released. Other flags include ULTRAPLAN for 30-minute planning sessions in the cloud, expanded VOICE_MODE, BRIDGE_MODE for bidirectional communication between IDE and CLI extensions, Undercover Mode with instructions to clean up AI traces in commits, and telemetry that tracks when you curse Claude as a frustration metric. Regarding risks: the permission logic and hooks are now public, and attackers can create malicious repositories. Hours before the leak, the axios package, a dependency of Claude Code, suffered a supply-chain attack with a Trojan in versions 1.14.1 and 0.30.4. It was updated via npm on March 31st between 00:21 and 03:29 UTC; check the lockfile. Anthropic recommends using a native installer instead of npm and rotating API keys as a precaution. The code is no longer reversible. At least 3 mirrors are already on GitHub. Claude Code is much more than a terminal-based model, and now anyone can study how it works. Anthropic is far ahead in engineering with 3-layer memory, KAIROS, and Coordinator Mode. But competitors now have the blueprint. 🔗 Sources: ▸ VentureBeat (Anthropic confirmation): https://venturebeat.com/technology/cl... ▸ Cybernews: https://cybernews.com/security/anthro... ▸ The AI ​​Corner (deep dive flags): https://www.the-ai-corner.com/p/claud... Join our Discord server and follow me on social media: 🟣 Discord Hora de Codar:   / discord   🔴 Instagram:   / horadecodar   TIMESTAMPS 00:00 Leak of all Claude Code code 01:30 All about the Anthropic leak Claude Code 03:30 The main leaked features of Claude Code 07:30 Anthropic (Claude Code) leak in the media 09:43 My thoughts on the Claude Code leaks