☀️ Anthropic Accidentally Shipped Its Own Source Code — Twice in One Week

# ☀️ Anthropic Accidentally Shipped Its Own Source Code — Twice in One Week

**April 1, 2026**

—

## 🔓 Claude Code’s Entire Codebase Leaked Via a Debugging File

Anthropic had a catastrophically bad Monday. Developers discovered that a routine npm package update for Claude Code (v2.1.88) included a source map file — a debugging artifact that pointed directly to a downloadable archive of the tool’s complete, unobfuscated TypeScript source code. All ~500,000 lines across ~1,900 files. Hosted on Anthropic’s own cloud storage. Anyone could grab it.

Within hours, the repo hit 84,000+ GitHub stars and tens of thousands of forks. Someone rewrote the entire thing in Python using OpenAI’s Codex overnight. Containment is impossible.

The leak revealed Claude Code’s internal architecture in remarkable detail: a sophisticated “harness” layer that governs memory, tool use, and autonomous behavior. Key finds include a feature called KAIROS that lets Claude Code run as a persistent background agent when users are idle, a self-healing memory system to prevent context overload, 35 build-time feature flags compiled out of public builds, and references to an unreleased model codenamed “Capybara.” Community members even used the leaked code to identify and patch a token drain bug affecting usage limits.

This is Anthropic’s second security blunder in a week, following a “Mythos” model leak days earlier.

**So what:** This is genuinely damaging for Anthropic on multiple levels. Claude Code is their fastest-growing product — 41% of professional developers use it daily, $2.5B in annualized revenue — and what makes it dominant isn’t just the underlying model, it’s the orchestration layer around it. That’s exactly what leaked. Every competitor (OpenAI, Google, startups) now has a detailed blueprint for how Anthropic solved problems like persistent agent memory, autonomous background execution, and context management at scale. For a company positioning itself as the “safety-first” AI lab preparing for an IPO, accidentally shipping your own source code twice in one week is the kind of trust problem that no blog post can fix. The irony is thick: the company that talks most about AI safety can’t secure its own npm packages.

Sources: [The Neuron](https://www.theneuron.ai/explainer-articles/anthropic-leaks-claude-code-a-literal-blueprint-for-ai-coding-agents/) · [CNBC](https://www.cnbc.com/2026/03/31/anthropic-leak-claude-code-internal-source.html) · [Ars Technica](https://arstechnica.com/ai/2026/03/entire-claude-code-cli-source-code-leaks-thanks-to-exposed-map-file/) · [The Rundown AI](https://www.therundown.ai/p/openai-new-122b-funding-superapp) · [Evolving AI](https://x.com/Fried_rice/status/2038894956459290963)

—

## 💰 OpenAI Closes $122B — the Largest Private Fundraise in History

OpenAI closed a $122 billion funding round at an $852 billion valuation. Amazon committed $50B, NVIDIA and SoftBank each put in $30B, and — for the first time — $3B came from individual retail investors through bank channels.

The numbers OpenAI shared are staggering: $2B in monthly revenue, 900M weekly active users, 50M subscribers. Their ads pilot reportedly hit $100M in annual revenue in under six weeks. Codex has 2M weekly users, up 5x in three months. Enterprise accounts for 40%+ of revenue and is growing fastest.

Buried in the announcement: they’re building a “unified AI superapp” merging ChatGPT, Codex, browsing, and agentic capabilities into one experience — which tracks with their recent shutdown of Sora as a standalone product.

**So what:** The enterprise revenue stat is the real story underneath the headline number. 40% of revenue from enterprise and climbing means OpenAI’s pivot away from consumer “side quests” (Sora, etc.) wasn’t a retreat — it was skating to where the puck was heading. The “superapp” framing matters too: OpenAI is betting that the future is one integrated AI experience, not a collection of point tools. For publishers and media companies, this means the platform you use for content creation, coding, research, and automation is increasingly going to be one thing — and whoever controls that interface controls the workflow. The retail investor tranche is also notable: this is the first time everyday investors could buy into a private AI round, which signals either peak hype or genuine democratization of AI wealth creation. Probably both.

Sources: [OpenAI Blog](https://openai.com/index/accelerating-the-next-phase-ai/) · [TechCrunch](https://techcrunch.com/2026/03/31/openai-not-yet-public-raises-3b-from-retail-investors-in-monster-122b-fund-raise/) · [The Rundown AI](https://www.therundown.ai/p/openai-new-122b-funding-superapp)

—

## 📊 Americans Are Using AI More — and Trusting It Less

A new Quinnipiac poll reveals a widening paradox: AI adoption is surging while public trust and optimism are sinking. Usage is up significantly, but Americans are increasingly uncomfortable with the pace of development, the companies driving it, and the lack of regulatory guardrails.

Separately, Jack Dorsey’s Block published a piece arguing that AI should replace the “coordination layer” of management entirely — essentially, that middle management is an information-routing problem that AI solves better than humans.

**So what:** This is the trust gap that should worry everyone in tech. People are using AI because they feel they have to — for work, for productivity, to keep up — not because they trust it. That’s adoption driven by anxiety, not enthusiasm, and it’s a fragile foundation. For news organizations specifically, this data point is critical: if public trust in AI is declining while AI-generated content proliferates, publishers who can credibly signal “humans made this” have a real differentiator. The Block piece about AI replacing middle management is provocative but reveals the cultural tension perfectly — the people building AI tools think management is a bug, while most workers see it as the last human layer between them and algorithmic control. Product teams should pay attention to both signals: build for AI-skeptical users, not AI evangelists.

Sources: [Quinnipiac University Poll](https://poll.qu.edu/poll-release?releaseid=3955) · [Block](https://block.xyz/inside/from-hierarchy-to-intelligence) · [Evolving AI Insights](https://evolvingai.io/subscribe)

—

## 📰 Judge Blocks Trump’s Defunding of NPR and PBS

A federal judge ruled that the Trump administration’s executive order to cut federal funding for NPR and PBS is unconstitutional, halting the defunding effort for now. The ruling is a significant win for public media, though appeals are expected.

**So what:** This matters beyond the immediate legal question. NPR and PBS serve as critical infrastructure for local news — hundreds of local public radio and TV stations depend on federal funding not just for their own operations, but as seed funding that unlocks additional grants and donations. If that federal floor disappears, the stations most likely to close are in rural and underserved areas that are already news deserts. The ruling buys time, but the political will to defund public media isn’t going away. For anyone who cares about the future of local journalism, this is a battle being fought in courts now, but it’ll be fought in budgets and ballot boxes for years.

Sources: [NYT](https://www.nytimes.com/2026/03/31/business/media/trump-npr-pbs-executive-order-ruling.html) · [Washington Post](https://www.washingtonpost.com/business/2026/03/31/trump-npr-pbs-funding-unconstitutional/) · [Poynter](https://mailchi.mp/poynter/how-a-cnn-reporter-covered-war-from-inside-iran)

—

## 🎙️ CNN’s Frederik Pleitgen on Covering War From Inside Iran

Poynter’s latest podcast features CNN’s Frederik Pleitgen describing what it’s like to report from inside Iran during an active war — the first Western journalist to do so. He and producer Claudia Otto traveled 24+ hours with no escort, navigating constant bombardment, restricted internet, government surveillance, and no safe zones. Unlike Ukraine, where you can retreat 20-30 miles behind front lines, Iran offered no such refuge.

An American journalist, Shelly Kittleson, was also kidnapped in Baghdad this week, underscoring the escalating dangers facing correspondents in the region.

**So what:** War reporting has always been dangerous, but the Iran conflict represents something different: a war zone with essentially zero Western press infrastructure, severe communications restrictions, and no established safety corridors. Pleitgen’s account matters because it shows what it actually costs — in human terms — to get ground-truth reporting from a conflict zone. At a time when AI-generated “coverage” and social media speculation can fill the information void in seconds, the gap between what an AI can synthesize from satellite imagery and what a reporter on the ground can witness is the gap between information and journalism. That gap is getting more expensive and more dangerous to bridge, which means fewer outlets will be willing to pay for it.

Sources: [Poynter Report Podcast](https://www.youtube.com/watch?v=UAKX1KCMxQI) · [NYT](https://www.nytimes.com/2026/03/31/world/middleeast/us-journalist-abducted-iraq-shelly-kittleson.html) · [Al-Monitor](https://www.al-monitor.com/originals/2026/03/american-journalist-shelly-kittleson-abducted-iraq)

—

## 🧪 Quick Hits

**🍎 Apple testing multi-command Siri** — Finally. Siri will process multiple requests in a single query instead of requiring individual asks. Expected at WWDC on June 8. Better late than never, but this was table stakes two years ago.

Sources: [TLDR](https://links.tldrnewsletter.com/Db0bgV)

**🔒 axios npm package compromised** — The axios library (300M+ weekly downloads) was hijacked via a compromised maintainer account and injected with malware. If your publishing stack uses JavaScript, check your dependencies.

Sources: [TechCrunch](https://techcrunch.com/2026/03/31/hacker-hijacks-axios-open-source-project-used-by-millions-to-push-malware/)

**🤏 1-bit AI model runs on iPhone** — PrismML launched Bonsai, an 8B-parameter model using 1-bit quantization that fits in 1.15 GB and runs at 40 tokens/sec on an iPhone. The trend toward capable, tiny models matters for on-device AI in newsrooms and publishing tools.

Sources: [The Neuron](https://theneuron.ai/explainer-articles/around-the-horn-digest-everything-that-happened-in-ai-today-monday-march-31-2026/)

**🌡️ AI data centers heating surrounding areas by up to 9.1°C** — Affecting 340 million people globally. The physical cost of AI infrastructure is becoming impossible to ignore.

Sources: [The Neuron](https://theneuron.ai/explainer-articles/around-the-horn-digest-everything-that-happened-in-ai-today-monday-march-31-2026/)

**📰 Salt Lake Tribune dropping its paywall in May** — The nonprofit news org is dismantling subscriber-only access entirely, betting on community support over gate-keeping. A bold experiment in sustainable local news.

Sources: Research brief community analysis

—

## 🎧 Podcast Queue

– **”The State of AI Q2: AI’s Second Moment”** — AI Daily Brief (23 min) — Quarterly report covering the agentic explosion, Claude Code’s revenue surge, and the “SaaS apocalypse”
– **”How to Ship Fast and Drive Outcomes”** — Hard Calls with Trisha Price (36 min) — Why product teams are still feature factories and what to do about it
– **”Mistral: Voxtral TTS, Forge, Leanstral & Mistral 4″** — Latent Space (49 min) — Deep dive on Europe’s most ambitious AI lab

—

*Compiled from: The Neuron, The Rundown AI, Semafor Flagship, Evolving AI Insights, TLDR, Poynter Report, overnight research brief, and podcast scan across 18 feeds.*