AI news April 3, 2026: $297B in VC, Anthropic leaks twice
Alexandre
··
Reading time: 15 min
If you thought last week's AI news was the peak with Cursor built on Kimi and the death of Sora, we just crossed a new threshold. This week, we're talking hundreds of billions, trillions of parameters, and 512,000 lines of code that were never supposed to see the light of day.
$297 billion in venture capital in a single quarter. OpenAI closing $122 billion, the largest funding round in tech history. Microsoft launching three foundational models to challenge Google and OpenAI on their own turf. Apple accidentally activating its AI features in China without regulator approval. And Anthropic suffering two major leaks in one week: first Claude Mythos, its secret 10-trillion-parameter model, then the full source code of Claude Code — 512,000 lines of TypeScript out in the wild. On X and Reddit, it was the historic verdict against Meta and YouTube that set discussions on fire: a jury declared for the first time that social networks are liable for their addictive design.
This is the kind of week where each announcement taken in isolation would be an event. For a solo dev building Livate with Claude Code — the tool whose source code just ended up on GitHub — this week is personal.
The week at a glance
Event
Date
Actor
Key figure
Global VC record Q1 2026
Apr. 1
Global market
$297 billion
Largest tech funding round ever
Mar. 31
OpenAI
$122B / $852B valuation
Launch of 3 MAI foundational models
Apr. 2
Microsoft
MAI-Image-2 at $5/M tokens
Accidental Apple Intelligence deployment
Mar. 30
Apple
~20% of iPhone market (China)
Claude Mythos leak
Late March
Anthropic
10 trillion parameters
Claude Code source code leak
Apr. 1
Anthropic
512,000 lines, 8,000+ DMCA
Addictive design verdict Meta/YouTube
Mar. 28
US courts
$6M damages ($4.2M Meta)
AI systemic risks warning
Apr. 2
Former Google/OpenAI/DeepMind leaders
Business Insider publication
Sources: TechCrunch, CNBC, Wall Street Journal, PCMag, SecurityWeek, NPR
$297 billion in Q1: venture capital enters uncharted territory
Global venture capital reached $297 billion in Q1 2026, according to TechCrunch. That's 2.5 times the previous quarterly record, and more than any full year before 2019. 41% of that amount — $122 billion — came from a single deal: OpenAI's funding round. Venture capital (VC) is investment in private companies in exchange for equity, with the expectation of a massive return at exit or IPO.
OpenAI closed its $122 billion round on March 31, valuing the company at $852 billion post-money (CNBC). Twelve months ago, OpenAI raised $40 billion at a $300 billion valuation. The valuation nearly tripled in a year. The Wall Street Journal calls it "the largest funding round in Silicon Valley history."
The surprising detail: OpenAI opened this round to retail investors, raising $3 billion from non-institutional participants (MLQ.ai). A first for an AI company of this size. The signal is clear: OpenAI is preparing its IPO and wants to build a retail shareholder base. On X, reactions swing between euphoria and vertigo. $852 billion for a company that's not yet profitable.
PitchBook notes that US VC funding alone hit $267 billion, dominated by OpenAI, Anthropic, and xAI. SiliconAngle confirms that AI is capturing a historically disproportionate share of available capital.
My take as a solo dev
What strikes me most is the concentration. Three companies capture the bulk of $297 billion. When a sector attracts this much capital this fast, two things happen: the tools get better (good for me), and the prices go up (less good). Every dollar flowing into OpenAI or Anthropic funds the models I use daily to build Livate. But this concentration also creates a structural dependency. If tomorrow Anthropic decides to triple the price of Claude Code, I don't really have an equivalent alternative.
These $297 billion don't change my day-to-day in the short term. But long-term, they determine who controls the tools I work with. And that's something I'm watching closely.
But while money floods into AI startups, Microsoft decided to play its own hand.
Click to enlarge
Microsoft launches 3 foundational models and takes on Google and OpenAI
On April 2, 2026, Microsoft launched three foundational models through its MAI Superintelligence team, led by Mustafa Suleyman, CEO of Microsoft AI. These include MAI-Image-2 for image generation, an audio transcription model, and a speech synthesis model, available on Microsoft Foundry and the MAI Playground, according to TechCrunch. A foundational model is a large-scale AI model trained on massive datasets that serves as a base for multiple specialized applications.
Pricing is the most interesting element. Microsoft is deliberately positioning itself between Google and OpenAI:
Model
Provider
Input price ($/M tokens)
Output image price ($/M tokens)
MAI-Image-2
Microsoft
$5
$33
DALL-E 3
OpenAI
~$4
~$40
Imagen 3
Google
~$3
~$12
Source: TechCrunch, April 2, 2026
What's strategically interesting is the timing. Microsoft invests massively in OpenAI — it's the largest shareholder — while simultaneously developing its own competing models. It's the same logic as Amazon distributing third-party products while launching its own. You fund your rival AND build the alternative. If it works, you win either way.
MAI Playground, launched March 19, lets developers test models before deploying them via Foundry. The integration with the Microsoft ecosystem (Azure, Office, GitHub) is the real competitive advantage: you go from prototype to production without switching platforms.
My take as a solo dev
Honestly, for an indie dev, the pricing war between Microsoft, Google, and OpenAI is the best thing that could happen. When three giants compete on the same segment, the cost per token drops. And the cost per token is what determines whether I can afford to let Claude Code iterate 50 times on a Livate component without checking the bill.
The less reassuring part is the ecosystem confusion. Microsoft is simultaneously an investor in OpenAI, a distributor of its models via Azure, and a direct competitor with MAI. For a dev choosing their stack, it makes the picture harder to read. Who's going to maintain what in 2 years? That's the kind of question that doesn't have a satisfying answer today.
Apple accidentally activates AI in China without regulator approval
On March 30, 2026, Apple inadvertently deployed Apple Intelligence on iPhones in China, even though the service had not received approval from the Cyberspace Administration of China (CAC), according to MacRumors and 9to5Mac. The CAC is the Chinese body that regulates all digital content and requires real-time filtering of AI models deployed in the country.
Bloomberg analyst Mark Gurman was the first to flag the anomaly: Apple would never launch its AI features on its largest foreign market without an official announcement, on a Sunday, and using Google's reverse image search — a service blocked in China (The Next Web). Apple quickly pulled the AI features from Chinese devices.
Apple's official plan for China relies on a partnership with Alibaba and its Qwen model, which must integrate a real-time filtering layer to meet CAC requirements. A separate agreement with Baidu covers the visual intelligence component. No relaunch date has been communicated.
Apple sells roughly 20% of its iPhones in China, representing over $60 billion per year. And CNBC reminds us that Apple is celebrating its 50th anniversary this week while trying to prove it can win the AI era. Accidentally activating an unapproved feature in the most regulatory-sensitive market in the world — tough timing.
My take as a solo dev
This is the kind of mistake where you can feel that the urgency to ship took over the validation process. Apple, the company known for obsessively controlling every last detail, accidentally deploys an unapproved feature in China. That means the pressure to catch up on AI is so intense that it's short-circuiting internal safeguards.
For devs building iOS apps, this is a warning sign. If Apple is juggling partnerships (Alibaba, Baidu) and regulations (CAC) for every market, the APIs you use might behave differently depending on the region. More fragmentation, more edge cases to handle. On Livate, I haven't integrated Apple Intelligence yet, but the day I do, this geographic complexity will need to be factored in from the design phase.
Anthropic's black week: Mythos leaks, then Claude Code's source code
In the span of one week, Anthropic suffered two major leaks. In late March, a draft blog post accidentally published on their CMS revealed the existence of Claude Mythos, a 10-trillion-parameter model described as a "step-change" in capabilities, according to Mashable. Then on April 1, a debug sourcemap accidentally included in Claude Code v2.1.88 exposed 512,000 lines of TypeScript across 1,900 files — the full source code of the agent, according to SecurityWeek and PCMag. A sourcemap is a technical file that maps compiled code back to its original source. Normally, you never ship it in production.
The timeline of the two leaks:
Date
Event
Immediate impact
~Mar. 28
Draft Anthropic blog post accidentally published on CMS
Reveals Claude Mythos (10T parameters, "unprecedented" cyber capabilities)
Apr. 1
Debug sourcemap in Claude Code v2.1.88
512,000 TypeScript lines exposed, recreated as "Claw Code" within 48h
Apr. 2-3
8,000+ DMCA takedown requests filed by Anthropic
Code still in the wild, prompt injection vulnerability identified
Sources: Mashable, SecurityWeek, PCMag
The Mythos details are staggering. The leaked draft describes a model capable of discovering and exploiting security vulnerabilities at scale — not just detecting them, exploiting them. Axios reports that Mythos's cyber capabilities are described as "far ahead of any other AI model." The leak also reveals the "Capybara" lineup, a series of models positioned between Mythos and the current Opus series. MediaPost notes that all of this comes at the worst possible time: Anthropic is preparing its IPO.
For Claude Code, the cascade was even faster. Within hours, researchers led by Sigrid Jin and Yeachan Heo reverse-engineered the code and created "Claw Code," a Python reimplementation that racked up tens of thousands of GitHub stars. A critical vulnerability in the command rate-limiting logic was identified, exploitable via prompt injection — the technique of injecting malicious instructions into an agent's context to hijack its behavior.
My take as a solo dev who uses Claude Code every day
I'll be straight: this week made me think. Claude Code is the tool I've been building Livate with for over a year. And its complete source code is out in the wild.
On one hand, it's almost reassuring. The code was scrutinized by thousands of developers within hours, and the architecture was generally praised. When your tool passes an involuntary community audit and comes out looking pretty good, that's a positive signal.
On the other hand, the prompt injection vulnerability identified in the aftermath is chilling. That's exactly the kind of flaw an attacker can exploit when they have the source code in front of them. For a dev who lets Claude Code access their file system, their API keys, their database, it's a brutal reminder: the tool is powerful, but so is the attack surface.
And Mythos at 10 trillion parameters is the promise that Claude will become even more capable. But two accidental leaks in one week from the company that positions itself as the most rigorous on safety — that's ironic. Anthropic builds the most advanced models on the market, but their own internal processes let a public CMS and a production sourcemap slip through. It's the kind of contradiction that reminds me that even the best make basic mistakes when the pressure ramps up.
Click to enlarge
AI systemic risks 2026: former Google, OpenAI, and DeepMind leaders sound the alarm
On April 2, 2026, former executives from Microsoft, Google, OpenAI, DeepMind, and the White House published a joint warning in Business Insider about AI systemic risks: amplification of inequality, large-scale automated cybercrime, massive job displacement, and concentration of power in a handful of organizations. A systemic risk, as opposed to a localized risk, is one that cascades through an entire interconnected system — like the 2008 financial crisis, but applied to AI.
What sets this warning apart from previous ones: the signatories' profiles. These aren't academics or activists. These are people who built these systems, saw from the inside how decisions are made, and are publicly stating that the current pace creates irreversible risks without coordinated governance.
CyberScoop reports that cybersecurity experts like Kevin Mandia (Mandiant) and Alex Stamos declared that the next two years are going to be "insane" in terms of automated cyberattacks. Axios confirms that AI is reshaping the cybersecurity landscape at a speed that outpaces defenders' ability to adapt.
GovTech adds a concept that went viral on X this week:
Definition — Superstupidity (Elon University, 2026): the greatest risk of AI isn't a malicious superintelligence, but massively deployed AI systems that make confidently wrong decisions at scale. Unlike superintelligence (a future and uncertain risk), "superstupidity" is a present and observable risk: agents that are highly confident in wrong answers, applied to millions of simultaneous decisions.
In parallel, McKinsey publishes its analysis of "The Great Flattening":
Definition — The Great Flattening (McKinsey, 2026): a term describing the elimination of middle management layers through the deployment of autonomous AI agents. According to Alexis Krivkovich, partner at McKinsey, this translates into wider spans of control, fewer hierarchical levels, and faster decision-making processes. The phenomenon is already affecting companies like Factory and IBM and is beginning to restructure the white-collar job market.
Forbes offers a nuance: the white-collar job "bust" will eventually turn into a "boom" once skills readjust. Asana's CEO anticipates this shift by positioning his platform as a "chaos orchestrator" for managing AI agents at work (Business Insider).
My take as a solo dev
The time lag is striking. The same week $297 billion flows into the sector, the people who built this industry say "careful, we're moving too fast." It's not a contradiction. It's the exact summary of where we are: everyone knows it's going too fast, and everyone keeps accelerating anyway.
For me, using Claude Code every day, the warning about automated cybercrime is concrete. Last week, I talked about the 40,000 exposed OpenClaw instances. This week, cybersecurity experts say the next two years will be "insane," and the source code of my main tool is out in the wild. On Livate, that's what I try to do with JWT, rate limiting, Helmet, parameterized queries via Sequelize. But I know the threat landscape is evolving faster than my defenses.
The Great Flattening — I live it every day. I'm a solo dev doing the work of 3-4 people thanks to code agents. What McKinsey describes for large enterprises is exactly what independent developers have been practicing for 18 months already. But it narrows the door for juniors. A senior with agents replacing a team of 5: efficient for the senior, catastrophic for the other 4. I'm the direct beneficiary of this trend. And I struggle to tell myself it's only good news.
Other AI news this week: Axios trojan, Meta convicted, Rebellions
News
What happened
Why it matters
Axios npm trojan
The npm account of the Axios maintainer — 300 million downloads, present in 70,000+ packages — was compromised on March 31. The attacker published version 1.14.1 with a malicious dependency (plain-crypto-js) containing a cross-platform RAT that self-destructed after execution and called its C2 within seconds. (Dark Reading, IT News)
If you run npm install without locking your versions, you're potentially affected. Dark Reading directly ties this incident to the week's source code leaks: when code is out in the wild, attackers know which dependencies to target.
Meta/YouTube verdict
LA jury: Instagram and YouTube found guilty of addictive design. $6M in damages ($4.2M Meta). Second jury: Meta fined $375M for failing to protect minors (NPR, CNN)
First time a jury accepts the "defective product" theory for an app. Went viral on Reddit r/legaltech. A legal precedent that goes beyond ethics.
Meta cuts 200 Bay Area jobs
519 total layoffs in California in 2026 (Mercury News)
Linked to AI agents in X/Reddit discussions. A concrete illustration of The Great Flattening.
Rebellions raises $400M
South Korean fabless AI chip startup, valued at $2.34B (Axios)
The race for NVIDIA alternatives intensifies in Asia.
iQIYI launches Nadou Pro
First Chinese AI agent for professional film/TV production. Trending on X (FT Markets)
A signal that agentic AI is going beyond code and into audiovisual production.
Steyer AI tax
Tom Steyer proposes taxing AI companies to fund retraining for displaced workers (Let's Data Science)
A political signal to watch in a pre-election year.
Data centers in space
Starcloud raises $170M at a $1.1B valuation. Tech billionaires advocate for orbital data centers to address the lack of terrestrial capacity (Payload Space)
A concrete answer to the energy/land bottleneck for data centers — or not.
What the week of April 3, 2026 changes for independent developers
1. Money is moving faster than technology
$297 billion in one quarter. Investors are betting that AI will transform every sector. For independent developers, that means better and cheaper tools in the short term, but also a growing dependency on platforms controlled by a few hyper-funded players.
2. The model war benefits users
Microsoft, Google, OpenAI, Anthropic: four giants competing on price and features. Cost per token drops, capabilities increase. This is the best time in history to be a dev using LLMs daily.
3. Leaks reveal how fragile processes become under pressure
Two leaks at Anthropic in one week, an accidental deployment at Apple: when the most rigorous companies in the sector let their secrets slip through basic errors, it's a sign that the AI race pressure is compromising internal controls.
4. Security is no longer optional, even for a side-project
512,000 lines of Claude Code out in the wild. Experts predicting two "insane" years of AI-powered cyberattacks. Machine-speed exploitable vulnerabilities. If you're shipping a product, you need to think security from day 1, not after the first incident.
5. The courts are entering the game
The Meta/YouTube verdict sets a precedent: tech platforms can be held liable for their product design. For devs building apps with engagement loops, this went from an ethical risk to a documented legal risk.
Frequently asked questions about AI news April 3, 2026
How much did OpenAI raise in Q1 2026?
OpenAI closed a $122 billion funding round on March 31, 2026, valuing the company at $852 billion post-money. It's the largest funding round in Silicon Valley history, according to the Wall Street Journal. For the first time, OpenAI opened this round to retail investors, raising $3 billion.
What is Claude Mythos, Anthropic's secret model?
Claude Mythos is an AI model developed by Anthropic whose existence was revealed through a leak in late March 2026. According to the leaked draft, Mythos would reach 10 trillion parameters and represent a "step-change" in capabilities, particularly in offensive cybersecurity — capable of discovering and exploiting vulnerabilities at a scale described as "unprecedented" compared to any other existing model.
How did Claude Code's source code end up on GitHub?
On April 1, 2026, Anthropic accidentally included a debug sourcemap in Claude Code v2.1.88, exposing 512,000 lines of TypeScript across 1,900 files. Within hours, developers reverse-engineered the code and created "Claw Code," a Python reimplementation. Anthropic filed over 8,000 DMCA takedown requests.
Why did Apple deploy AI in China by accident?
On March 30, 2026, Apple inadvertently activated Apple Intelligence in China without CAC regulator approval. The incident was flagged by Bloomberg analyst Mark Gurman. Apple urgently pulled the AI features. The official plan relies on a partnership with Alibaba (Qwen model) and Baidu to meet Chinese regulatory requirements.
What is "The Great Flattening" according to McKinsey?
The Great Flattening is a term introduced by McKinsey in 2026 to describe the elimination of middle management layers through the deployment of autonomous AI agents. Companies like Factory and IBM are widening their spans of control, reducing hierarchy, and accelerating decisions — at the cost of fewer middle management positions.
Conclusion: the week the masks came off
$297 billion. $122 billion for a single company. 512,000 lines of source code out in the wild. Apple Intelligence activated by mistake on hundreds of millions of iPhones. And the people who built all of this saying "we're moving too fast."
For a solo dev building with Claude Code every day, this week has a particular flavor. The tool I use daily passed an involuntary community audit — and came out looking pretty solid. That's the paradox: Anthropic's leaks revealed the quality of the code just as much as the fragility of the processes.
The question is no longer "is AI changing everything." It's "do you trust the people building it."
Alex
Key takeaways
$297B in VC during Q1 2026, an all-time record. OpenAI closes $122B at an $852B valuation — the largest funding round in Silicon Valley history. Retail investors are invited for the first time.
Microsoft launches 3 foundational models via MAI Superintelligence (Mustafa Suleyman). MAI-Image-2 at $5/M tokens input, positioned between Google Imagen and DALL-E.
Apple accidentally deploys Apple Intelligence in China on March 30 without CAC regulator approval. The feature is pulled urgently. The Alibaba/Baidu partnership remains on hold.
Anthropic suffers two leaks: Claude Mythos (10 trillion parameters, 'unprecedented' cyber capabilities) via a public CMS, then 512,000 lines of Claude Code source code via a sourcemap. The community recreates the agent as 'Claw Code'. 8,000+ DMCA requests filed.
Former Google, OpenAI, and DeepMind leaders publish a joint warning on systemic risks. Historic verdict: Meta and YouTube found liable for addictive design ($6M in damages). McKinsey theorizes 'The Great Flattening' driven by AI agents.
