A missing .npmignore exposed 512,000 lines of Claude Code's internal TypeScript source
Leaked codenames include Mythos (next-gen model), Capybara (new tier above Opus), and KAIROS (always-on background agent)
Opus 4.7 and Sonnet 4.8 are incremental updates, Mythos is the real generational leap
Anthropic confirmed Mythos exists and called it "the most capable model we've built to date"
An AI design tool and Claude Builder for full-stack app creation could drop alongside Opus 4.7 this week
Someone at Anthropic forgot to add two characters to a config file. Those two characters were *.map in a .npmignore. The result was the largest accidental source code exposure in AI history. On March 31, 2026, version 2.1.88 of the Claude Code npm package shipped with full source maps intact, dumping 512,000 lines of internal TypeScript across 1,906 files into the public registry. Within hours, the entire codebase was archived on GitHub and dissected by thousands of developers.
This is the story of what they found, why it matters, and what it tells us about where Anthropic is headed.
The Leak Nobody Saw Coming
The root cause was almost comically simple. Bun, the JavaScript runtime Anthropic uses to build Claude Code, generates source maps by default during compilation. Source maps are debug files that map compiled code back to original source. They are useful in development and catastrophic in production when they contain your entire proprietary codebase.
Standard practice is to exclude these files from published packages using .npmignore. Someone on the Claude Code team missed that step. The .map file went out with the package, and suddenly every npm install @anthropic-ai/claude-code came bundled with the full source tree.
The timing made it worse. Just five days earlier, on March 26, roughly 3,000 internal documents had leaked through what appeared to be a CMS misconfiguration. Those files referenced an unreleased model codenamed Mythos and described it as a restricted security research tool. Then the npm leak dropped and confirmed everything the CMS files hinted at, plus a lot more.
What the Source Code Revealed
The 512,000 lines contained version strings, feature flags, internal architecture, and codenames that paint a clear picture of Anthropic's roadmap.
44 hidden feature flags. These control unreleased capabilities that are already built but not yet exposed to users. The flags suggest features ranging from enhanced memory systems to new collaboration modes.
Model codenames. The source referenced several internal model names: Fennec (the current Opus 4.6), Numbat (an unreleased model), and most notably, Capybara. Capybara appears to be a new model family sitting above the current Opus tier. Cross-referencing with the earlier CMS leak, Capybara is almost certainly the model Anthropic internally calls Mythos.
KAIROS. This was the most surprising find. KAIROS is an always-on background agent embedded directly into Claude Code. According to the source, it runs persistent monitoring tasks, tracks context across sessions, and performs maintenance operations without explicit user commands. The name suggests it handles time-sensitive operations (kairos is Greek for "the opportune moment"). Whether KAIROS is active in current builds or gated behind a feature flag remains unclear.
Stealth mode. The source included a mechanism designed to hide Anthropic employee contributions to open-source projects. The feature masks authorship metadata so that commits from Anthropic engineers appear to come from generic accounts. This raised eyebrows in the open-source community, though Anthropic has not commented on its purpose.
Mythos vs. Opus 4.7: Two Very Different Things
The most common misunderstanding in coverage of these leaks is conflating Opus 4.7 with Mythos. They are not the same thing.
Opus 4.7 is the next incremental update within the existing Claude 4.x model family. Anthropic has been shipping these roughly every three to four months. Opus 4.5 arrived last November, Opus 4.6 in early February 2026, and Opus 4.7 is expected this week. These are version bumps. Better benchmarks, improved reasoning, faster inference. Important, but evolutionary.
Mythos is something else entirely. Anthropic's own description, confirmed through the leaked documents and a preview page on their security research portal, calls it "a step change and the most capable model we have built to date." Mythos is not a point release. It is a generational leap, the kind of model that redefines what the system can do.
The leaked source also references Sonnet 4.8, which follows the same incremental pattern as the Opus line. So the picture looks like this: the 4.x family continues to evolve with regular updates (Opus 4.7, Sonnet 4.8), while Mythos represents a parallel track, a next-generation architecture being developed and tested separately.
Currently, Mythos access is restricted to select security research partners. Anthropic is using it for vulnerability research, including zero-day discovery. Twelve founding partners reportedly have access. This limited rollout suggests the model is powerful enough that Anthropic wants to understand its risk profile before wider release.
The Design Tool Nobody Expected
Alongside Opus 4.7, Anthropic is preparing an AI design tool that generates complete websites and presentation decks from natural language prompts. This is not a minor feature addition. It signals Anthropic's move beyond chat and code into full-stack creative production.
The tool reportedly combines content creation, visual design, and technical implementation in a single workflow. You describe what you want. It builds the page, writes the copy, handles the styling, and outputs production-ready code.
A separate feature called Claude Builder takes this further. It provides a template-based interface for creating full-stack applications with real-time previews, integrated security measures, and deployment tooling. Think of it as Lovable or Bolt, but backed by the same model that powers Claude Code.
Anthropic has also partnered with Figma to convert AI-generated code back into editable design files and integrated Claude into Microsoft Word and PowerPoint. The strategy is clear: Anthropic is not just building a chatbot or a coding assistant. It is building an end-to-end creative and technical production platform.
What This Means for Developers
The practical takeaway splits into short-term and long-term.
Short-term: Opus 4.7 drops this week. If you are on Claude Code, expect better reasoning, faster responses, and likely some of those 44 feature flags getting switched on. The model upgrade will be automatic for API users and Claude Code subscribers. Watch for the design tool launch alongside it.
Long-term: Mythos changes the calculus. A model that Anthropic describes as a step change, powerful enough to discover zero-day vulnerabilities, is going to reshape expectations for what AI systems can do autonomously. When Mythos eventually gets broader access, the gap between "AI assistant" and "AI colleague" gets a lot smaller.
The KAIROS revelation is equally significant. An always-on background agent means Claude Code is not just responding to your commands. It is actively monitoring, maintaining, and optimizing your workflow in the background. If this feature ships publicly (and the source code suggests it will), the interaction model for AI coding tools changes fundamentally. You stop telling it what to do and start collaborating with something that already knows what needs doing.
The Irony
The biggest AI coding company in the world got caught by a missing line in a config file. The same tool that helps thousands of developers catch exactly this kind of mistake shipped with the mistake itself. Anthropic's response was swift: they pulled the affected package version within hours and issued a security advisory. But the code was already archived, forked, and analyzed.
There is a lesson here that goes beyond Anthropic. Source maps, build artifacts, debug logs. These are the things that slip through when shipping velocity outpaces operational hygiene. It happens to everyone. It just happened to happen to the company whose product is literally designed to prevent it.
The leak was embarrassing. What it revealed was fascinating. And what comes next, Opus 4.7 this week, Mythos on the horizon, a design tool that turns prompts into products, that is the part worth paying attention to.
Top comments (0)