While most people were watching AI from the sidelines, something unprecedented happened — and the developers, builders, and thinkers who understood it first are already three steps ahead.
On March 31, 2026, security researcher Chaofan Shou noticed something that shook the AI industry: the complete source code of Claude Code — Anthropic's flagship AI coding agent generating $2.5 billion in annualized revenue — was sitting on the public npm registry in plain sight. A 59.8 MB JavaScript source map file, accidentally bundled into version 2.1.88 of the @anthropic-ai/claude-code package, pointed directly to a ZIP archive containing 512,000 lines of unobfuscated TypeScript across 1,906 files. Within hours, it had been mirrored across GitHub, amassing tens of thousands of stars. Within 48 hours, developer sentiment toward Anthropic had completely reversed — from skepticism to awe.
If you're part of the community that follows AI seriously, what happened next is the story that matters.
What Is Claude Code and Why This Leak Is a Big Deal
Claude Code is Anthropic's AI-powered terminal agent — a coding assistant that doesn't just suggest code but executes tasks, manages files, calls APIs, and operates with increasing autonomy directly on your computer. Launched publicly in May 2025, it has become the fastest-growing developer tool in Anthropic's portfolio, generating $2.5 billion in annualized recurring revenue as of February 2026, with enterprise clients accounting for 80% of that total.
The leaked code is not Claude's model weights — it's the "agentic harness" that sits around the AI: the orchestration logic, the tool-call loops, the permission architecture, the memory system. According to Fortune and VentureBeat, this is the software layer that gives Claude Code its behavior, its guardrails, and its ability to act autonomously. What competitors got was not the brain — they got the blueprint for building the body.
Anthropic confirmed the incident immediately. A spokesperson told CNBC:
Earlier today, a Claude Code release included some internal source code. No sensitive customer data or credentials were involved or exposed. This was a release packaging issue caused by human error, not a security breach. We're rolling out measures to prevent this from happening again.
Boris Cherny, a Claude Code engineer at Anthropic, added on social media:
Mistakes happen. As a team, the important thing is to recognize it's never an individual's fault. It's the process, the culture, or the infra.
What Was Hidden Inside — The Features No One Was Supposed to See Yet
While the majority of developers saw a headline about a leak, those who went deep found something far more interesting: a detailed roadmap of where agentic AI is actually going.
KAIROS — The Always-On Autonomous Agent
Hidden behind a feature flag, KAIROS is a fully built background agent that runs 24/7 without user prompting. Every few seconds it receives a heartbeat: "anything worth doing right now?" If it acts, it can fix errors, respond to messages, update files, and run tasks entirely on its own — with push notifications to your phone or desktop even when the terminal is closed. This is not a chatbot. This is an AI that works while you sleep.
ULTRAPLAN — Frontier-Level Planning
ULTRAPLAN offloads complex planning tasks to a remote Opus 4.6 session that gets up to 30 minutes to think through a problem, then teleports the completed plan back to your local terminal. Multi-Agent Coordinator Mode extends this further, turning Claude Code into an orchestrator that spawns parallel worker agents, assigns tasks, and synthesizes results.
Dream Mode — Background Memory Consolidation
A background "Dream" subagent periodically runs through four phases — Orient → Gather → Consolidate → Prune — maintaining a self-healing memory system that prevents the "context entropy" problem: the tendency of AI agents to become confused or hallucinatory as sessions grow longer. The architecture uses a lightweight MEMORY.md index of pointers (~150 characters per line) that lives permanently in context, with actual knowledge stored in topic files fetched on demand.
Buddy — The Companion System
Buddy is a Tamagotchi-style terminal companion with 18 species, rarity tiers, a 1% shiny rate, procedurally generated RPG stats, and ASCII art sprites. Your buddy is deterministically assigned based on your user ID hash. The launch window coded into the source: April 1–7, 2026. Someone at Anthropic is building personality into the product — intentionally engineering user attachment.
Undercover Mode
A subsystem designed to auto-activate for Anthropic employees working on public repositories, stripping AI attribution from commit messages. The irony: Anthropic built an AI-powered leak prevention system, then accidentally published 512,000 lines of proprietary code.
The Numbers That Define the Scope
- 512,000 — lines of TypeScript exposed in the leak.
- 1,906 — files contained in the source map archive.
- $2.5 billion — annualized ARR of Claude Code as of February 2026.
- $19 billion — Anthropic's total annualized revenue run rate as of March 2026.
- 80% — share of Claude Code revenue from enterprise clients.
- 84,000+ — GitHub stars accumulated by leaked mirrors within days.
- Hundreds — of public repository mirrors, despite DMCA takedowns from Anthropic.
This was the second major leak in one week. Days earlier, Fortune had reported that Anthropic accidentally published a draft blog post revealing an upcoming model codenamed both "Mythos" and "Capybara" — confirmed as a Claude 4.6 variant with capabilities described as unprecedented in cybersecurity applications.
Why This Is a Signal, Not Just a Scandal — And What It Means for You
Here's what the majority missed while debating whether this was embarrassing for Anthropic: the leak confirmed that the most sophisticated agentic AI systems in the world are already built. They are not coming. They are sitting behind feature flags, waiting for the right moment to ship.
If you're a developer, this is your inside look at the architecture that will define the next wave of AI tools. The three-layer memory system, the KAIROS heartbeat model, the multi-agent coordinator — these are the patterns that serious builders will be reverse-engineering and implementing in their own stacks over the next 12 months.
If you're a founder or product leader, the revelation that 80% of Claude Code's $2.5 billion ARR comes from enterprise clients tells you exactly where AI agent monetization is concentrating. The companies building on top of these architectures — not those debating whether to — are the ones capturing that value.
If you're a professional watching AI from the outside: the agents revealed in this leak are not futuristic. They are in production, behind flags, at a company with $19 billion in annualized revenue. The gap between those who understand what's being built and those who don't is closing fast. The people in the room who read this kind of coverage don't wait to be told when to move. They've already moved.
What Comes Next — The Optimist's Reading
The leak won't sink Anthropic. Claude Code's $2.5 billion ARR is built on enterprise contracts and trust, not secrecy. Within 48 hours of the exposure, developer sentiment shifted from frustration to excitement — because what was exposed showed a team building at a level that most competitors can't yet match.
The real lesson the industry takes from this is not about Anthropic's operational security. It's about the trajectory of agentic AI. KAIROS, ULTRAPLAN, Dream mode, and multi-agent coordination aren't experiments. They're nearly-shipped features in a product already generating billions in revenue. That means the AI agent era isn't a prediction. It's a release schedule.
Want to follow every major AI move the moment it happens? Join thousands of readers at fabioraphaelli.com and get the biggest AI news every week.
Frequently Asked Questions
What exactly was leaked in the Anthropic Claude Code incident?
On March 31, 2026, Anthropic accidentally published a 59.8 MB JavaScript source map file in the @anthropic-ai/claude-code npm package version 2.1.88. The file contained 512,000 lines of unobfuscated TypeScript across 1,906 files — the complete client-side agentic harness of Claude Code, including unreleased features, memory architecture, orchestration logic, and internal feature flags. No model weights, customer data, or credentials were exposed.
What is KAIROS and why are developers excited about it?
KAIROS is a fully built but unreleased always-on background agent hidden behind a feature flag in Claude Code's source. It runs autonomously 24/7, receiving a heartbeat prompt every few seconds asking "anything worth doing right now?" It can fix errors, update files, run tasks, and send push notifications to your phone without you initiating anything. It represents the most advanced publicly visible implementation of autonomous AI agent architecture.
Did the Claude Code leak harm users?
Anthropic confirmed no customer data, credentials, or model weights were exposed. However, security firm Zscaler warned that the leak coincided with a separate malicious supply chain attack on the Axios npm package on the same day. Users who updated Claude Code via npm between 00:21 and 03:29 UTC on March 31, 2026, should audit their lockfiles for Axios versions 1.14.1 or 0.30.4 and the dependency plain-crypto-js.
What is Anthropic's "Buddy" companion system?
Buddy is a Tamagotchi-style terminal companion hidden in the Claude Code source under buddy/companion.ts. It features 18 species, rarity tiers, a 1% shiny rate, procedurally generated RPG stats including CHAOS and SNARK levels, and ASCII art sprites. Each user's Buddy is deterministically assigned based on their user ID hash. The planned launch window in the source code was April 1–7, 2026.
What does the Claude Code leak tell us about the future of AI agents?
The leak confirms that always-on autonomous agents (KAIROS), multi-step planning systems (ULTRAPLAN), background memory consolidation (Dream mode), and multi-agent orchestration are fully built features at Anthropic — not research projects. With Claude Code already generating $2.5 billion in annualized revenue and 80% enterprise adoption, these features represent the near-term roadmap for production AI agents, not theoretical futures.









