Story Highlights May 1st 2026

Pentagon cleared seven AI companies for classified networks — Anthropic's still locked out for saying no. Musk admitted in court he trained Grok on OpenAI data. And China just made AI-driven layoffs illegal. Let's get into it.

The Pentagon signed classified AI contracts with seven companies — and Anthropic, the one that refused, remains barred.

On May 1st, the Defense Department announced agreements with Nvidia, Microsoft, Amazon Web Services, Reflection AI, Google, OpenAI, and SpaceX. All seven are now authorized to deploy their models on Impact Level 6 and IL7 classified networks — the DoD's highest security tiers — for what officials called "lawful operational use." Over 1.3 million DoD personnel already use GenAI.mil, the Pentagon's AI platform, which officials say is cutting tasks from months down to days. The deals follow Anthropic's refusal to give the military unrestricted access to its models — specifically its objection to use in autonomous weapons and domestic mass surveillance. The Pentagon responded by branding Anthropic a "supply-chain risk," a designation normally reserved for foreign adversaries. A California judge granted an injunction against that label in March; the D.C. Circuit allowed the designation to stand while litigation continues.

Every major American AI lab is now inside the Pentagon's classified networks except the one that publicly drew a line. Pentagon CTO Emil Michael called Anthropic's situation a "separate national security moment" on Friday. Trump signaled a possible thaw last week, saying Anthropic was "shaping up." But for now, seven companies hold military contracts and Anthropic has a lawsuit. The move also reflects the Pentagon's stated goal of preventing AI vendor lock-in — a strategy that's clearly working faster than Anthropic anticipated.

The man who co-founded one of those seven companies spent his third straight day on the federal stand this week.

The Musk v. Altman trial is three days in — and it's already produced a courtroom moment even the judge didn't see coming.

Elon Musk is suing OpenAI, CEO Sam Altman, president Greg Brockman, and investor Microsoft for over $130 billion. He wants OpenAI reverted to a nonprofit structure and both executives removed. The trial began April 28th in Oakland. Musk co-founded OpenAI in 2015, gave at least $44 million, and left after a power struggle in 2018. He argues OpenAI's 2025 conversion to a public benefit corporation betrayed its founding mission; OpenAI says Musk wanted total control and sued when he couldn't get it. Day 3 went sideways: Musk's finance manager Jared Birchall accidentally testified about xAI's $97.4 billion bid for OpenAI assets — evidence that had been blocked from discovery. Judge Gonzalez Rogers dismissed the jury and interrogated Birchall directly. "Sounds like you wanted to open the door," she told Musk's lawyers.

If Musk wins, OpenAI's IPO plans collapse and a forced return to nonprofit status would cut off its capital pipeline — making this one of the highest-stakes AI cases in history. The Birchall slip may open the door to antitrust scrutiny of xAI's own conduct. The trial has also surfaced decade-old emails revealing Musk largely drafted OpenAI's original mission statement. What started as a grudge lawsuit is beginning to look like a genuine fight over who controls the trajectory of AI.

Meanwhile, Apple just confirmed it had no idea how many Macs people actually wanted.

Apple's Mac division posted a quarter it didn't expect — powered by the same AI wave everyone else is scrambling to build for.

Apple's Q2 2026 earnings showed Mac revenue of $8.4 billion, beating Wall Street expectations of roughly $8 billion and up 6% year-over-year. Total company revenue hit $111.2 billion, up 17%. CEO Tim Cook told analysts that demand for Mac mini and Mac Studio exceeded Apple's own forecasts — both models have sold out — driven by users running local AI models, particularly OpenClaw, on Apple silicon. The MacBook Neo, launched March 4th, is also supply-constrained. Cook said it may take "several months" to restore balance. School systems including Kansas City Public Schools are switching from Chromebooks to the Neo. Enterprise companies like Perplexity have adopted Mac as their platform for building AI assistants. The Mac mini is now the top-selling desktop in China, where OpenClaw demand has been intense.

This reframes what a Mac is. It's no longer just a productivity device — it's becoming a local AI server. Apple's silicon is competitive enough that consumers and enterprises are choosing on-device processing over cloud dependency. That's good for Apple's hardware margins but puts pressure on a supply chain the company underplanned. Competitors are now on notice: hardware that can run powerful local models at consumer prices is suddenly the battleground.

Anthropic's newest model is showing what AI looks like when the capabilities get genuinely alarming.

Anthropic released a model that found zero-day vulnerabilities in every major operating system and browser — and deliberately chose not to release it to the public.

On April 7th, Anthropic unveiled Claude Mythos Preview, a general-purpose model whose cybersecurity capabilities emerged as a byproduct of improved coding and reasoning, not by design. The model found a 27-year-old vulnerability in OpenBSD, a 17-year-old remote code execution flaw in FreeBSD, a 16-year-old bug in FFmpeg, and multiple Linux kernel privilege escalation chains. In UK government testing, Mythos achieved 73% on expert-level capture-the-flag challenges and became the first model to complete a 32-step simulated corporate network attack from start to finish. Anthropic is keeping it gated through Project Glasswing — a coalition including AWS, Apple, Microsoft, Google, CrowdStrike, and Nvidia — plus 40 additional organizations. Anthropic committed $100 million in usage credits and $4 million to open-source security organizations.

Finance ministers, central bankers, the Bank of England, EU regulators, and IMF participants have all raised alarms. Pentagon CTO Emil Michael called Mythos a "separate national security moment." The UK's AI Safety Institute concluded the model could successfully attack poorly defended systems, while hardened environments would likely hold. The danger window everyone is worried about: the gap between when AI finds a vulnerability and when humans can patch it.

Staying in the courtroom — Musk made an admission on Thursday that complicated his own lawsuit considerably.

On the stand, Elon Musk confirmed that xAI trained Grok using OpenAI's models — while actively suing OpenAI for betraying its mission.

During cross-examination, OpenAI attorney William Savitt asked Musk whether xAI had used model distillation on OpenAI's outputs to train Grok. Musk said "generally all the AI companies" do this. Pressed for a yes or no, he said "partly." He added that using other AI models to validate your own is "standard practice." Distillation trains a smaller model on the outputs of a larger, more capable one — letting competitors acquire capabilities at a fraction of the cost and time. Later in testimony, Musk ranked the AI leaders: Anthropic first, then OpenAI, Google, and Chinese open source models. He described xAI as "much smaller," with just a few hundred employees.

The irony is hard to overstate. Musk is demanding $130 billion from a company whose work his own AI apparently benefited from. The admission also confirms what many in the industry assumed: even as U.S. labs publicly condemn Chinese distillation, the practice runs in all directions. Distillation may not be clearly illegal, but it significantly weakens the moral framing Musk has built his lawsuit around.

Google reported this week that AI agents browsing the web are already under active attack.

Google scanned billions of web pages and found malicious instructions targeting AI agents embedded across the public web right now.

On April 23rd, Google's threat intelligence team published findings from a sweep of Common Crawl — a database of billions of crawled public web pages. They found real indirect prompt injection payloads: hidden instructions buried in HTML, metadata, and white space, designed to hijack AI agents that browse those pages. Unlike jailbreaks, these attacks hide inside content the AI is explicitly told to read. Categories range from harmless pranks and SEO manipulation to commands for data exfiltration and file deletion. Most are low sophistication. Security researcher Ron Stoner demonstrated the threat independently: for $12 and a single Wikipedia edit, he convinced multiple AI chatbots he was the reigning world champion of a card game called 6 Nimmt! — a competition that doesn't exist.

As AI agents gain real permissions — read emails, manage calendars, execute transactions — this attack surface becomes genuinely dangerous. A poisoned webpage could trigger an agent to send money, leak credentials, or delete files, and it would look identical to normal operation. Defenses exist: isolated sanitizer models to strip untrusted content, zero-trust agent permission structures, and forensic audit trails on every AI decision. The current threat is experimental. The future threat, as agents get more capable, is not.

China resolved a significant labor dispute this week by ruling that AI is not an acceptable reason to fire anyone.

Chinese courts ruled companies cannot legally terminate employees just because an AI can do the job cheaper.

The Hangzhou Intermediate People's Court upheld a ruling for worker Zhou, a quality assurance supervisor whose company replaced his role with an AI model, then offered him a reassignment at 40% less pay. When he refused, they fired him. The court ruled that AI replacement does not qualify as a "major change in objective circumstances" under Chinese labor law — the legal threshold that permits termination. A similar ruling in Beijing in December 2025 reached the same conclusion for a data mapping worker replaced by AI. Courts stated clearly that AI integration should "liberate labor and promote jobs," and that companies must offer retraining, fair reassignment, or compensation — not dismissal. China's core AI industry exceeded 1.2 trillion yuan in 2025, and the government expects AI penetration above 90% by 2030.

This sets a significant precedent in the world's largest tech employment market. China is simultaneously pushing aggressive AI adoption and protecting workers from its consequences — a tension most governments haven't resolved. The ruling signals that the cost of automation cannot simply be transferred to employees. As Western AI adoption accelerates with far weaker labor protections, this week's ruling may look prescient.

Every country and every company is choosing which side of AI they stand on.