You are building on AI. Your team uses Claude, GPT, or Gemini in their daily work. The pipeline works, automations run, clients are happy. And then, in one week, a company that spends billions on AI safety and employs the best engineers in the world accidentally publishes 3,000 internal documents and half a million lines of source code. This is not a hypothetical scenario. This Anthropic leak happened in the last week of March 2026. And the lessons learned from it are more relevant to your company than you might think.
Anthropic Leak: Two Incidents, Five Days
One email with a concrete point each week.
AI, B2B sales, and implementations. No spam, unsubscribe with one click.
Mythos: Secret AI Model (March 26)
Anthropic left publicly available around 3,000 unpublished resources on its website. Draft blog posts, internal materials, images, PDFs. The reason? A CMS configuration error that caused draft documents to end up in a publicly indexed data store. It was first reported by Fortune magazine.
Among the documents was a description of a new AI model codenamed Mythos (internal project name: Capybara).
What do we know from this Anthropic leak about Mythos:
- „This is a new class of model, larger and more intelligent than Opus models.”
- Anthropic calls it „breakthrough change” (step change) in AI capabilities
- Programming, reasoning, and cybersecurity test results are significantly better than Claude Opus 4.6
- The internal assessment states plainly: Mythos is „far ahead of every other AI model in cyber capabilities.”
The most serious thread? Documents show that Anthropic privately warns US government officials before the rise of large-scale cyberattacks in 2026. More about the threats Euronews described.
But that's not the end of this week.
Claude Code source code on npm (March 31)
4:23 AM Eastern Time. Chaofan Shou, an intern at Solayer Labs, is downloading a routine update Claude Code z npm registry. The package weighs 59.8 MB. Normally, it's a few megabytes. He opens it, looks, and sees half a million lines of Anthropic's complete source code. Feature flags, system prompts, product plans. Everything. Before Anthropic could react, the code was already on GitHub. Deleting the package from npm didn't help.
How did this happen? Claude Code uses the Bun bundler, which generates source maps by default. A source map file is a debugging tool that maps compressed code back to readable source. All it took was a missing entry in .npmignore. A detailed analysis of this Anthropic leak dev.to.
Scale? 1,900 files. 512,000+ lines of TypeScript code. Full, readable code for Anthropic's flagship AI programming tool.
What was found after the Anthropic leak in the code?
Code analysis revealed things that Anthropic would prefer to keep to themselves. And which say a lot about where the entire industry is heading.
44 feature flags
Code found 44 feature flags. Switches controlling access to built, but inactive features. These are not prototypes. This is ready, compiled code waiting to be turned on.
KAIROS: Agent operating non-stop
The biggest discovery from the Anthropic leak. KAIROS (from the Greek for „right time”) appears in the code over 150 times. It turns Claude Code from an on-demand tool into an autonomous background daemon. GitHub webhook sessions, cron job scheduling, automatic resume without commands. And the most interesting part? The process autoDream, which is „memory consolidation” during times when the user is inactive. AI that works while you sleep.
But KAIROS is not the only discovery. The next one is even stranger.
BUDDY: virtual pet
No one expected this. Claude Code has a system virtual Tamagotchi-style companion. Rarity-varying species, shiny variants, procedurally generated stats. Buddy sits in a speech bubble next to the text field. Yes, seriously.
Coordinator Mode and others
In the catalog Coordinator hidden is a multi-agent orchestration system. It turns Claude Code into a coordinator managing multiple agents in parallel. In addition to this: voice command mode, browser control via Playwright, persistent memory between sessions, internal employee-only tools, and built-in system prompts.
Anthropic Leak: Company Response
The company confirmed the existence of Mythos and described it as a „new class of model” representing a „breakthrough change.” Due to cybersecurity concerns, it is limiting early access to cybersecurity defense organizations.
Regarding the Claude Code, the advocate said:
„Today's Claude Code release contained some internal source code. No customer data or credentials were exposed. This was a release packaging issue caused by human error, not a security breach.”
The package disappeared from npm. Too late. The code was already on GitHub.
Anthropic Leak and Your Company's Security
Open Door Syndrome
Most companies think a breach is a hacking problem. That someone has to break in. The truth is, the most dangerous breaches are misconfigurations. A missing entry in .npmignore. Publiczny bucket S3. Draft w CMS-ie bez flagi „prywatny”. Zero hakerstwa, 100% ludzkiego błędu.
Anthropic. Billions in budget. Top-tier engineers. The industry's best AI safety practices. And all it took was a missing entry in a single configuration file.
No one broke in. The door was just open. That's open door syndromeCompanies invest in locks, alarms, and surveillance, and then leave keys under the doormat. We previously wrote about securing AI systems in a company.. These rules are now more important than ever.
Questions for your team
You might be thinking, „But we're not Anthropic. We don't publish AI models or npm packages.” Fine, but honestly ask yourself: do you know exactly what's in your Docker image? Has anyone checked what the CMS is publishing publicly? Are API keys or test data definitely not lurking in the repository?
Some specific questions for the next standup:
- Does a CI/CD pipeline check what actually ends up in the production package?
- Is anyone scanning artifacts before publication?
- Are you monitoring package sizes? 59.8 MB is a warning sign.
- Are the files
.map,.envAre test data excluded from production?
Feature flags reveal the strategy
44 Flags is Anthropic's complete product roadmap. Features the company has never publicly discussed. For OpenAI And Google is priceless knowledge.
If your company uses feature flags, ensure that flags are removed from the code after the feature is enabled. Make sure their names don't reveal business plans. And that the code behind them doesn't end up in public artifacts.
Mythos changes the rules of the game
Documents indicate that AI agents can lead multiple hacking campaigns simultaneously. That employees using AI agents may unknowingly open the door to criminals. That identity theft is becoming easier than ever. Anthropic warns the US government that this is a real threat for 2026.
Summary: Lessons from the Anthropic Leak
| What happened | Lesson |
|---|---|
| 3,000 internal documents publicly | Regularly audit what is available on your infrastructure |
| Source map with full code in npm | Automate artifact checks in CI/CD |
| 44 hidden features revealing plans | Remove flags from production code |
| Client-side system prompts | Keep sensitive instructions on the server side |
| Model with cyber abilities | Before you check the locks, check if the doors are closed |
What to do after the Anthropic leak?
Good news? Each of these problems can be solved. None of the protections we describe here require months of work or a million-dollar budget. These are changes your team can implement this week.
- Review CI/CD pipeline. Check what goes into the production packages
- Audit public resources. CMS, repositories, package registries, S3 buckets
- Update the AI security policy The risks of AI agents and new models
- Does your team know what a source map is? Train them.
- Monitor artifact sizes. 59 MB CLI package is a red flag
The code for Claude Code is already on GitHub. Hackers are analyzing it, looking for attack vectors on companies that use this tool. Mythos, a model with „unprecedented cyber capabilities,” could be available in weeks.
If your team is using AI tools in your daily work, now is the time to check if your pipelines, artifacts, and configurations are secure. We perform code audits and optimization (so-called Web Vitals). A 30-minute call, a concrete checklist to make your app more secure or achieve better SEO results without bogging down your users. Schedule a free consultation.
