Claude Mythos Leaks, Apple Turns 50, Iran’s Cyber War, and the $1B Seed Round

Networkcraft Desk  ·  March 30, 2026

📊 This Week in Numbers

The Leak: Claude Mythos Is Real, and Anthropic Says It’s Scary Good

A CMS misconfiguration at Anthropic this week accidentally exposed approximately 3,000 unpublished internal files — and hidden among them was something the AI industry had only whispered about: Claude Capybara / Mythos, described internally as “by far the most powerful model we have ever built” and representing “a genuine step change in capabilities.” The disclosure was not planned. It was a mistake. And it may be one of the most consequential accidental leaks in AI history.

The files were discovered by Roy Paz, a researcher at LayerX Security, and Alexandre Pauwels, a PhD candidate at Cambridge University. They flagged the exposure to Anthropic within hours, but not before the contents had been cached and indexed. The leaked assets included internal capability benchmarks, training notes, and deployment strategy documents — all pointing to a model with autonomous offensive cybersecurity capabilities that could independently find and exploit software vulnerabilities without human guidance.

Anthropic’s response was striking. Rather than simply deploying the model commercially, the company confirmed it plans to restrict Claude Mythos’s initial rollout exclusively to cyber defenders — security teams, threat researchers, and red-teamers — precisely because of its ability to autonomously identify and chain vulnerabilities. This is the first time a major AI lab has pre-emptively limited a model’s release on capability-safety grounds before it’s even publicly announced.

What does this mean for the broader capability arms race? It means Anthropic has drawn a line — but also confirmed they crossed it. The existence of Mythos tells us that frontier labs are now building models they themselves consider potentially dangerous. The debate about how to handle that is no longer theoretical.

Apple at 50: Mac Pro Dies, Siri Gets Gemini, WWDC June 8

Apple turns 50 on April 1, 2026. And to mark the occasion, the company is doing what it does best: quietly killing beloved hardware and doubling down on AI. The Mac Pro has been officially discontinued — a product line that spanned 20 years and defined professional computing for a generation of creative workers. Its replacement? The Mac Studio, which Apple argues now offers equal or greater performance in a fraction of the footprint. Whether that’s an engineering triumph or a marketing reframe depends on who you ask.

On the software side, iOS 27 is shaping up to be the biggest Siri overhaul since the assistant launched in 2011. Apple is building a full standalone Siri chatbot application and, in a move that would have seemed unthinkable three years ago, partnering with Google to use Gemini as the underlying model layer for complex queries. This comes alongside an “Extensions” marketplace that will allow Claude, Gemini, and Grok to all be accessible directly from within Siri — essentially turning Apple’s platform into a neutral AI switchboard.

WWDC 2026 has been confirmed for June 8–12, and hardware news is already flowing ahead of the developer conference. AirPods Max 2 pre-orders opened this week at $549, featuring improved noise cancellation and a redesigned mesh headband. Meanwhile, AirPods Pro 3 briefly hit an all-time low of $199 in a promotional window that sent Apple deal communities into a frenzy. At 50, Apple is leaning hard into the AI era — but it’s doing so by opening the platform rather than trying to own every layer.

OpenAI: $100M Ads in 6 Weeks, Sora Killed, $120B Raised

OpenAI’s ad business, launched just six weeks ago as a quiet pilot, is already printing money. With over 600 advertisers participating and 85% of users now eligible to see sponsored content, the company’s annualized ad revenue is already tracking above $100 million. That’s a faster ramp than almost any ad business in tech history, including early Facebook and Google. The secret? ChatGPT’s extraordinary user intent signal — people telling an AI exactly what they’re trying to do creates ad targeting opportunities that traditional search can’t replicate.

Less triumphant was the quiet death of Sora, OpenAI’s text-to-video model, which was shut down this week — just three months after a marquee partnership with Disney was announced to significant fanfare. The reason cited internally: “unsustainable compute economics.” Generating high-quality video at scale requires orders of magnitude more compute than generating text, and OpenAI apparently decided the unit economics simply didn’t work at their current infrastructure cost. The Disney deal, reportedly, has been renegotiated.

On the capital side, OpenAI’s rolling fundraise has now crossed $120 billion in total committed capital, with Amazon, Nvidia, and SoftBank among the most recent participants. The company is nearly doubling its workforce — from 4,500 employees to a target of 8,000 by end of 2026 — and has opened engineering offices in three new cities. The revenue line that justifies all of this: somewhere between $5B and $8B annualized, depending on which source you trust.

Iran’s Cyber-Physical War: Spyware in Bomb Shelter Texts

The most disturbing story of the week received the least mainstream coverage. Iranian cyber operations have reached a new threshold of sophistication and brutality: during missile strikes on Israeli population centers this month, SMS spyware was deployed to civilian phones at the exact minute the rockets hit — targeting people who were frantically searching for bomb shelter locations. The malware was designed to harvest device contents, location data, and contact networks under cover of the chaos.

The Handala group — Iran’s most aggressive hacktivist-aligned operation — took credit for wiping computers at defense contractor Stryker in real time during a coordinated cyber-physical strike sequence. More alarming still: the FBI Director’s personal devices were reportedly compromised in a separate operation, marking the first confirmed breach of a sitting U.S. intelligence chief since the OPM hack of 2015.

DigiCert’s threat intelligence division tracked approximately 50 distinct Iran-linked cyber groups responsible for 5,800+ attacks since February 28, 2026 — the day the Iran War formally began. Targets span U.S. critical infrastructure, Israeli hospitals, Gulf state government networks, and Western financial institutions. AI is central to the offensive: automated attack generation, AI-voiced phishing calls, and deepfake disinformation videos that have collectively amassed over 100 million views on social platforms.

This is no longer cyber espionage. It is cyber warfare with kinetic synchronization — and the scale and AI integration represent a structural shift in what nation-state attacks look like in 2026.

The Biggest VC Week in History

If you work in venture capital and weren’t busy this week, you weren’t paying attention. AMI Labs raised a $1.03 billion seed round — the largest seed round ever raised in European history, and arguably the most philosophically interesting funding event of the AI era. The company is founded by Yann LeCun, Turing Award winner and Meta’s Chief AI Scientist, who has spent three years publicly arguing that large language models are a fundamentally flawed path to intelligence. AMI Labs is building on his JEPA (Joint Embedding Predictive Architecture) framework — world models that learn by predicting structured representations of the world rather than predicting the next token. This is LeCun betting a billion dollars that he’s right.

Nscale raised $2 billion in a Series C at a $14.6 billion valuation — the largest equity round in European tech history — with Goldman Sachs and JPMorgan acting as placement agents. Replit closed $400 million at $9 billion, approaching $1 billion in ARR. And the deals kept coming: Legora ($550M), Mind Robotics ($500M), Rhoda AI ($450M), and Nexthop AI ($500M) all closed this week.

The macro picture: Q1 2026 VC investment is now on pace to exceed the entirety of 2024 investment — a year that itself set post-correction records. This isn’t a bubble in the traditional sense; the companies being funded have real revenue and real customers. But the valuations are pricing in scenarios that require everything to go right.

Google Pixel March Drop + Android 16

Google’s monthly hardware drop cadence continued this week with a significant update to Circle to Search, which now includes a “Find the Look” feature that lets users circle any object in a photo — a piece of furniture, a clothing item, an architectural detail — and instantly surface visually similar products across Google Shopping and the wider web. It’s one of those features that sounds minor until you use it and realize you’ve been waiting for it for years.

Pixel Watch received safety upgrades including improved fall detection with automatic emergency services contact and a new “check-in” feature that pings a designated contact if a workout session goes silent for more than 20 minutes. And Android 16 began its Pixel 8+ rollout this week — the first major version to ship before Google I/O in years, reflecting the company’s accelerated OS release cycle.

The broader strategic question here is interesting: Google is now shipping meaningful hardware and software updates every 30 days, while Apple still anchors to its annual WWDC-plus-September rhythm. Monthly drops create more ongoing media coverage and keep the Pixel in headlines year-round. The tradeoff is that each individual feature gets less attention and testing time. Whether that matters to consumers is still an open question — but the two companies have clearly chosen opposite philosophies.

SpaceX IPO & Shield AI: Space + Defense Reach the Public Markets

SpaceX’s long-awaited IPO is no longer a “someday” story. The company is actively targeting a $1.75 trillion valuation for its public offering, with a $75 billion primary raise that would make it the largest IPO in history. What’s genuinely unprecedented: SpaceX is allocating 30% of shares directly to retail investors — a figure that dwarfs the typical 5–10% retail allocation in major tech IPOs. The move is partly Musk being Musk, partly a recognition that SpaceX has a retail fan base that would generate enormous demand, and partly a political statement about democratizing access to transformational companies.

Shield AI closed a $2 billion raise at a $12.7 billion valuation, with Advent International leading and JPMorgan and Blackstone providing an additional $750 million in debt facilities. Shield AI builds autonomous military AI pilots — systems that can fly fighter jets, coordinate drone swarms, and execute tactical missions without human pilots. The fact that Blackstone, a blue-chip PE firm, is now providing debt financing to a defense AI company is a significant signal: this sector has graduated from niche venture darling to institutional-grade asset class.

For retail investors watching this space: SpaceX’s 30% allocation means real access, but a $1.75T valuation bakes in a lot of optimism about Starship commercial success, Starlink subscriber growth, and the defense contracts pipeline. Do your homework before the hype machine hits full velocity.

Two things happened this week that we’ll look back on. First, Anthropic confirmed it has an AI model so capable it thinks it’s dangerous — and its response was to give it to hackers. That’s either brilliant or terrifying depending on your threat model. Handing an autonomous vulnerability-exploitation system to the people defending the internet is logical. It’s also a remarkable acknowledgment that the technology has crossed a threshold the company itself finds concerning.

Second, Yann LeCun raised a billion dollars to bet against the very architecture that made every AI headline this year. One of these people is right about where intelligence comes from. We don’t know which one yet. But by the time we find out, the capital will have already moved.

— Networkcraft Desk, March 30, 2026

🔢 The Number of the Week

5,800

The number of cyberattacks launched by Iran-linked groups since February 28, 2026 — the day the Iran War began. That’s 85 attacks per day, every day, for five weeks. And those are just the ones DigiCert tracked.

🔗 Related Reading

• → Weekly Brief #011: OpenAI’s Ad Business Launches, Meta’s Llama 4, and the Defense AI Boom
• → Weekly Brief #010: Gemini 2.5 Pro, Tesla FSD v14, and the $100B AI Infrastructure Sprint
• → AI Safety at Frontier Labs in 2026: What “Restricted Deployment” Actually Means

❓ Frequently Asked Questions

📧 Never Miss a Brief

The Weekly Brief lands every Monday morning.

Tech news, AI developments, cybersecurity, and the capital markets moves that matter — curated and written by the Networkcraft Desk. No noise. No fluff.

Subscribe Free →