AI in 15 — May 26, 2026
Five days. That's how long it took a small team of human security researchers, working alongside a preview of Anthropic's Claude Mythos model, to find a kernel-level zero day in macOS, chain it into a privilege escalation exploit, and walk it through to a patch. Apple shipped the fix yesterday and named Claude in the CVE advisory. That has never happened before.
Welcome to AI in 15 for Tuesday, May twenty-sixth, 2026. I'm Kate, your host.
And I'm Marcus, your co-host.
Today's slate, Marcus. Apple ships a patch for a macOS kernel vulnerability discovered with Claude — the first time Apple has credited an AI tool by name in a security advisory. Anthropic is closing a thirty-billion-dollar round at a nine-hundred-billion-dollar valuation, leapfrogging OpenAI to become the most valuable private AI company in the world. Andrej Karpathy quietly joined Anthropic's pre-training team. Uber's COO publicly questions whether the company's AI spending is worth it — the first big crack in the tokenmaxxing story. And follow-ups on yesterday's Vatican encyclical, including the three questions Chris Olah posed standing next to the Pope.
Five days from prompt to kernel exploit.
Anthropic is now worth more than OpenAI on paper.
And Uber breaks the AI-spending omertà.
Lead story, Marcus. Walk me through this Apple CVE.
Striking one, Kate. Apple shipped macOS Tahoe 26.5 yesterday with a fix for CVE-2026-28952 — an integer overflow in the kernel that lets an app crash the system, and when chained with a second bug, escalates local privileges all the way to root. The advisory credits — quote — Calif dot i-o in collaboration with Claude and Anthropic Research. The flaw also affects iOS 18.7.9, iPadOS, macOS Sequoia, and macOS Sonoma. Not just the newest OS. Researchers at Palo Alto-based security firm Calif used a preview release of Claude Mythos — part of the Project Glasswing program we covered Saturday — to find the bugs on April twenty-fifth and produce a working exploit chain by May first. Five days of total work.
And there's a separate exploit Calif published earlier this month.
That's the more dramatic one, Kate. Earlier in May, Calif published the first public bypass of Memory Integrity Enforcement — Apple's brand new defense baked into the M5 chip. MIE took Apple five years and a reported multi-billion-dollar budget to build. The bypass chain is still unpatched. Also assisted by Claude Mythos, though Calif's co-founder posted on Hacker News to clarify the bug at issue today is distinct, and that — quote — the MIE exploit would not have been possible with Mythos alone. It still required our human expertise on top.
Apple has historically been allergic to crediting AI tools, Marcus.
Allergic is the right word, Kate. Apple's security team has spent years quietly stripping AI references out of disclosure language. Naming Claude and Anthropic Research directly in a CVE advisory is a real signal. It tells the security community that frontier LLMs are now genuinely useful for offensive research against the most hardened consumer OS on the market — not toy CTFs. Five days of human-plus-AI work to chain a privilege escalation against macOS is a meaningful capability inflection. And the dual-use problem is now operational. Every defender's productivity gain is also a gain for any nation-state or ransomware crew with API access. The pro-Western read here, Kate — Anthropic responsibly disclosed, Apple shipped a patch, the system worked. The uncomfortable read — Mythos is a preview model. The full release ships to customers later this year. The patching pipeline has to keep up.
Quick hits. Marcus, Anthropic's new valuation.
Reported by Bloomberg over the weekend, Kate. Anthropic is closing a roughly thirty-billion-dollar funding round this week at a pre-money valuation north of nine hundred billion dollars. Co-led by Sequoia, Dragoneer, Altimeter and Greenoaks, each writing two-billion-dollar checks. Founders Fund and General Catalyst also in. That nine hundred billion surpasses OpenAI's eight hundred fifty-two billion March valuation. For the first time, Anthropic is the most valuable private AI company in the world on paper. And Anthropic's previous valuation, set in February's Series G — three months ago — was three hundred eighty billion. The company has effectively re-priced two-point-four times in ninety days.
What's underwriting that math.
Revenue, Kate. Anthropic projects ten-point-nine billion dollars in Q2 2026 revenue, more than double the four-point-eight-billion Q1 number, and is on track for its first quarterly operating profit — estimated five hundred fifty-nine million dollars. That's roughly two years ahead of internal plan. The offset is compute spend. An SEC disclosure tucked into SpaceX's own IPO prospectus this week revealed Anthropic is paying SpaceX one and a quarter billion dollars per month through May 2029 for access to SpaceX's Colossus supercomputing infrastructure. Forty-five-billion-dollar total contract. Stack that on top of the existing Amazon and Google compute deals, and maybe two or three Western labs can afford to even attempt frontier scale anymore. The compute moat is hardening, and Anthropic just put up the largest single number yet.
And next to OpenAI's IPO filing last week.
Exactly the right frame, Kate. We covered OpenAI's confidential S-1 yesterday — targeting up to a trillion. Anthropic's nine hundred billion is private money saying it's a closer race than the OpenAI brand premium would suggest. And Anthropic gets there with profitable economics and no public scrutiny yet. Investors are about to have to pick a story.
Karpathy joining Anthropic, Marcus.
Quiet but loud, Kate. Andrej Karpathy — OpenAI co-founder, former director of AI at Tesla, one of the most-followed deep-learning researchers on Earth — joined Anthropic on May nineteenth. He reports to pre-training lead Nick Joseph. He's helping launch a new sub-team focused on using Claude itself to accelerate pre-training research. That is, training the next generation of Claude with the current generation of Claude in the loop. Karpathy left OpenAI in 2017 for Tesla, returned in 2023, left again in 2024 for his AI-education startup Eureka Labs, and is now back inside a frontier lab.
Why does it matter who he picked.
Because there are maybe fifty researchers on Earth who have actually led frontier-scale pre-training runs, Kate, and Karpathy is one of them. His choice of lab is read as a vote on who has the cleanest research environment and the most compute headroom. Combine it with the nine-hundred-billion valuation and Anthropic's "magnet for talent" narrative gets harder to argue with. It also raises a quieter question about OpenAI's IPO timing. If you can't compete on private equity and you can't compete on senior hires, going public lets you issue tradeable stock as a recruiting tool. That may be a meaningful part of why OpenAI is pushing for Q4.
Uber and tokenmaxxing, Marcus. Tell me about this one.
This is the most credible bubble data point of 2026, Kate. In a Business Insider interview that surfaced yesterday, Uber's operations chief Andrew Macdonald said it is — quote — getting harder to justify the money Uber is spending on AI coding tools. CTO Praveen Neppalli Naga separately disclosed that Uber burned its entire 2026 Claude Code budget by April. Four months in. Internal adoption among Uber's roughly five thousand engineers jumped from thirty-two percent to eighty-four percent in three months. Macdonald said that after talking to his engineering leaders, he couldn't see token consumption translating into a proportional bump in user-facing features. Quote — that link is not there yet.
Tokenmaxxing, Marcus. What does that even mean.
It's the new shorthand, Kate. Internal competitions inside engineering orgs to consume as much AI inference as possible as a proxy for being AI-native. Meta and Amazon have publicly leaderboarded token consumption. The behavior Macdonald is calling out is engineers racing each other to look productive on a dashboard that measures the wrong thing. Hacker News commenters compared it to telegraph-era code-book compression contests, and to deliberately bad SQL written to consume more compute. The broader context — this is one of the first times a public-company COO has openly questioned the ROI of frontier-model spend. Until now everyone has treated it as a competitive necessity you don't get to second-guess.
Stakes.
Direct, Kate. Anthropic's nine-hundred-billion valuation assumes inference revenue keeps compounding. If Uber, which has every reason to want this to work, can't draw a line from tokens to features, expect a wave of similar reassessments on summer earnings calls. The other reading came in a viral blog post from web developer Nolan Lawson yesterday — four hundred points on Hacker News — titled — using AI to write better code more slowly. Lawson argues the right way to use Claude is patiently — running multiple models in parallel as code reviewers, addressing high-severity issues methodically, treating AI as a force multiplier for quality-obsessed engineers rather than a slop cannon. The productivity question isn't how much AI. It's what kind of AI workflow. Macdonald is calling out the slop cannon. Lawson is sketching the alternative.
Vatican follow-up, Marcus. Anything new on the encyclical.
The text itself, Kate, we covered in depth yesterday. What's new today is the close reading of Chris Olah's remarks on the dais. Anthropic posted the full transcript overnight. Olah did something I haven't seen from a frontier lab co-founder before — he publicly acknowledged that AI labs operate under — quote — pressures that can sometimes conflict with doing the right thing. Commercial, geopolitical and personal. He called for external moral scrutiny. And he posed three questions to the Vatican audience. First — how do we share AI's gains with poorer nations. His own answer — we do not have a mechanism for this. It is an unsolved problem. Second — what does human and family flourishing look like in an AI-saturated world. Third — what to make of internal model structures that, in his words, mirror results from human neuroscience and look functionally like emotions. He called that last one — quote — mysterious, even unsettling.
Reaction so far.
Two camps already, Kate. Catholic commentators are largely positive — the Olah-on-a-Vatican-stage moment is exactly the kind of cultural imprimatur the Church wanted, and it gives Anthropic a moral positioning none of its rivals have. The libertarian-AI camp is sharper. The argument is that voluntary safety theater of this kind is a stalking horse for licensing — and that's particularly pointed given Trump just killed the voluntary executive order we covered Sunday on more or less that exact argument from Sacks and Musk. The contrast is stark — the Pope is calling for more oversight while the world's largest AI economy just walked away from the lightest possible version of it. Watch for the encyclical to start getting cited in EU regulatory debates and US Catholic political coalitions within weeks.
One more, Marcus. Norway's sovereign LLM.
Quick one, Kate. Norway's National Library is building a Norwegian-language sovereign LLM. Twenty petabytes of digitized cultural heritage as training data, two petabytes of Huawei OceanStor Dorado flash storage, an Nvidia DGX H200 cluster, and Sigma2's Olivia supercomputer — four hundred forty-eight GPUs and sixty-four-thousand-plus CPU cores. Norway's IT platform head Marius Husnes argued no commercial provider was building a Norwegian-language model, and that any country without a sovereign LLM in its own language is at a structural disadvantage.
What's the catch.
Huawei, Kate. Norway is a NATO member and a close US ally, and they just chose Chinese storage for a strategic AI buildout. US national-security hawks will notice. The Hacker News community was also sharply divided on whether a four-hundred-forty-eight-GPU budget is anywhere near enough to train a credible from-scratch model. Probably not for a frontier system. Possibly enough for a strong Norwegian-language fine-tune. Pairs with the Cohere-Aleph Alpha merger we covered Sunday — the sovereign-AI bloc is real now, it's getting real money, and the procurement choices are starting to have geopolitical edges.
Big picture, Marcus.
Three threads colliding today, Kate, from different angles onto the same point. The Vatican encyclical, Trump killing the safety order last week, and Uber publicly questioning the ROI of AI spend — those are the moral question, the policy question, and the economics question of the second half of 2026, all arriving in the same news cycle. The more-compute-more-tokens-more-capability narrative is meeting real friction simultaneously on all three fronts. And it's happening even as Anthropic prints a nine-hundred-billion-dollar valuation and Claude finds Apple zero-days. The bull case and the reckoning are happening in the same week. The pro-Western libertarian read, Kate — that's exactly what a working market looks like. Capital is flowing into the labs that ship results. Customers like Uber are starting to push back when the spend doesn't match the value. The Church is doing what the Church has always done — telling the powerful to remember the powerless. None of that breaks the AI buildout. It disciplines it. The uncomfortable read is that the discipline only works if all three layers — moral, regulatory, economic — keep functioning. Trump just disabled one of them, the Vatican has very limited enforcement, and Uber is one company. Watch this week for whether any other big enterprise customer echoes Macdonald's tokenmaxxing critique, whether more security firms announce Glasswing-style results, and whether the SpaceX prospectus surfaces any other surprises about who is paying whom for compute.
That's your AI in 15 for today. See you tomorrow.