Google I/O's Real Story Is the Data Trade: Gemini Spark Wants Your Inbox, Your Files, and Your Trust — and the Comment Section Says You're Out. A Jury Dismisses Every Musk v. Altman Charge on the Statute of Limitations While Musk Was on Air Force One With Trump. Andrej Karpathy Leaves His Education Startup to Join Anthropic R&D. And Researchers Used Claude to Crack Apple's Most Hardened macOS Security Feature in Five Days.
The day after Google I/O is always the morning when the keynote's polish gives way to harder questions — and this year's hardest question arrived in the comment section of a Verge piece titled "Google's AI future demands trust — and your personal data." The answer from the public: no, and not anymore. Gemini Spark, Google's new always-on AI agent that runs 24/7 in the background monitoring your Gmail, scanning your credit card statements, and interacting with your local files, is the clearest expression yet of what Google needs from users to make its AI platform thesis work. The Musk v. Altman trial ended yesterday afternoon after roughly two hours of jury deliberation — not with a verdict on the merits, but a unanimous dismissal on statute of limitations grounds while Musk himself was aboard Air Force One bound for Beijing with President Trump. Andrej Karpathy announced he will join Anthropic for R&D, returning to the frontier AI world he helped build after a detour through an education startup. And a security research team published findings that researchers used Claude to build working exploit code for two macOS vulnerabilities in five days — cracking Apple's Memory Integrity Enforcement, a security feature Apple spent half a decade building.
The Morning After Google I/O: Gemini Spark Is Google's Always-On AI Agent That Runs 24/7 in Your Background, Monitors Your Inbox, Scans Your Credit Card Statements, and Can Access Files on Your Mac. Google AI Ultra Drops From $249 to $100 a Month. Gemini 3.5 Flash Ships as the Default Model Across All Google Products. And the Real Debate Isn't the Feature List — It's Whether Anyone Will Opt In.
The headlines from Google I/O 2026 ran predictably through the tech press yesterday: Android XR glasses, Googlebooks, Gemini Intelligence as the Android operating layer. But the story that is actually defining the day after I/O is a quieter one, surfaced in the comment section of a Verge piece that asked the obvious question behind every I/O announcement: what does Google get in exchange for all of this? The answer is unambiguous. Gemini Spark — Google's answer to OpenClaw, Claude, and every other always-on AI agent that has gained traction in 2025 and 2026 — runs in the background continuously using virtual machines on Google Cloud, connecting to Gmail, Google Docs, Sheets, Slides, and an expanding list of third-party services via the Model Context Protocol. This summer, Spark will also access local files on Mac computers. The pitch is that you barely have to interact with it: you toss things over your shoulder and Spark catches them, completing the task while you move on with your day. Concretely: it can write emails to your dog's boarder using vaccination records from your local drive, generate updated study guides from your course materials, and scan monthly credit card statements for subscriptions you have forgotten about.
Gemini Spark is powered by Gemini 3.5 Flash — announced at I/O as Google's strongest model yet for agentic and long-horizon tasks — which is now the default model across the Gemini app, Google AI Mode, and all Google consumer products. Gemini 3.5 Pro is expected in June. Alongside Spark, Google announced Gmail Live, a voice mode built into Gmail's search bar that lets you ask natural-language questions about your inbox and get spoken answers sourced directly from your email history, with visual citations so you can verify what you're hearing. Docs Live and a revamped Keep with voice-driven note-taking round out the Workspace AI update. Google AI Ultra dropped from $249.99 a month to $100, with a $200 tier that includes access to Project Genie, Google's world-model for interactive AI-generated environments. The new Antigravity desktop app — the previously teased AI development hub — launched alongside an updated command-line interface and an SDK for developers building their own agents on top of Google infrastructure. And Google's AI Studio vibe-coding tool is coming to Android, letting mobile users prompt and build apps on the go.
The thread running through all of it is access. Every feature that is genuinely useful in the I/O lineup requires Google to know more about you: what is in your inbox, what files are on your computer, what subscriptions you pay for, what is on your calendar. Josh Woodward, VP of Google Labs, told reporters that "millions of people are using [Personal Intelligence] every single day" — which is a strong data point about consumer willingness to hand over personal data for useful AI features. But the comment sections tell a different story: "Google lost the first [my trust] long ago and will never get any more of the second [my personal data] out of me" was the day's top-voted response on the Verge piece covering Google's I/O AI pitch. The gap between Google's product vision and consumer trust is not a new problem — but I/O 2026 has made it undeniable that the entire Gemini platform thesis depends on closing that gap. Spark does not work as a $100-a-month value proposition if users decline to opt in to the data connections that make it useful. Google's challenge is not building the features. It built them. The challenge is convincing people that the company whose advertising business was built on behavioral data has earned the kind of intimate, always-on access to your personal life that it is now asking for.
theverge.com ↗The Musk v. Altman Trial Is Over. A Unanimous Jury Dismissed Every Charge on Statute of Limitations Grounds — Not on the Merits. The Judge Accepted the Verdict. Musk Was on Air Force One With Trump Bound for Beijing When It Was Announced and Posted That He Would Appeal. OpenAI's Response: "This Lawsuit Has Always Been a Baseless and Jealous Bid to Derail a Competitor."
The tech trial of the year ended yesterday afternoon with something considerably less dramatic than either side had promised it would be. After roughly two hours of deliberation — an unusually short period given the three weeks of testimony, the 150-to-200 "I don't recall" instances counted by OpenAI's lead counsel William Savitt, and the procedural chaos that preceded it — the advisory jury returned a unanimous verdict dismissing all charges brought by Elon Musk against Sam Altman, Greg Brockman, and Microsoft. The two core claims — breach of charitable trust and restitution — were barred by the statute of limitations. The third claim, that Microsoft aided and abetted a breach of charitable trust, failed as a consequence. Judge Yvonne Gonzalez Rogers accepted the advisory verdict and dismissed the case. Musk was not in the courtroom when the verdict was announced. He had been absent since his own testimony, and by the time the verdict dropped he was aboard Air Force One with President Trump, en route to Beijing, and posted his response on X from the air.
Musk's statement on X: "The judge and jury ruled on a calendar technicality rather than the merits of the case. There is no question to anyone following the case in detail that Altman and Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!" He announced he would appeal. OpenAI did not issue an immediate public statement but had previously characterized the lawsuit as "a baseless and jealous bid to derail a competitor" in a statement posted to its newsroom X account during an earlier phase of the proceedings. Microsoft spokesperson Alex Haurek said the company welcomed the decision and called the facts and timeline "long clear." The trial had featured testimony from Musk, Altman, Brockman, Microsoft CEO Satya Nadella, former chief scientist Ilya Sutskever, and former OpenAI board member Shivon Zilis — and produced, as a side effect, the most detailed and unflattering portrait of the interpersonal dynamics among AI's founding generation that has ever been made public under oath.
The legal outcome is straightforward. The cultural outcome is more complicated. A trial where the central accusation was that OpenAI abandoned its founding mission to benefit humanity in favor of enriching its leadership produced testimony suggesting that the founding generation of the AI industry was, as a group, more motivated by rivalry, personal grievance, and competitive positioning than by the idealistic framing of their public narratives. Musk came out looking petty and unprepared. Altman came out looking polished but evasive. Mira Murati testified that she couldn't trust Altman's words. The jackass trophy — an award Josh Achiam received for being yelled at by Elon Musk — almost entered evidence. The statute of limitations dismissal means the central factual question — whether Altman and Brockman did what Musk accused them of — was never adjudicated. It will not be adjudicated in any appeal. The appeal will be about the limitations ruling. What we are left with is three weeks of testimony that illuminated how the AI industry actually operates at the level of its most consequential relationships, without any of it being weighed against a legal standard of truth. Everyone got to say damaging things about everyone else, nothing was conclusively resolved, and Musk has announced he will keep litigating. This is probably not the last time a courtroom will be asked to sort out the founding mythology of the AI industry. It is just the first time, and the first time was messy enough to make you wonder what subsequent attempts will look like.
theverge.com ↗Andrej Karpathy Is Joining Anthropic. The Former Tesla AI Director and OpenAI Co-Founder Who Has Been Running an Education Startup Is Returning to Frontier AI Research. He Says He Will Work on R&D, Is Still "Deeply Passionate About Education," and Plans to Return to It "In Time." The Move Adds One of AI's Most Respected Technical Voices to the Company Competing Most Directly With OpenAI.
Andrej Karpathy announced on X yesterday that he will be joining Anthropic to work on research and development. The announcement was brief and characteristic of Karpathy's public communication style: he said he would be working on R&D at Anthropic, that he remained "deeply passionate about education," and that he planned to return to that work "in time." He did not specify what he would be working on at Anthropic, what his role title would be, or how the arrangement would interact with his ongoing education project — an AI-native school concept he announced in 2024 that has been in development since he left Tesla. Anthropic did not immediately issue a statement beyond confirming the hire.
The career arc here is worth tracing precisely because of what it says about where the people who built this industry think the action is. Karpathy was a founding team member at OpenAI in its original nonprofit configuration, left to become Tesla's director of AI and Autopilot Vision for several years, returned to OpenAI for a second stint, and then in mid-2024 left again to pursue an "AI-native school" — a project he described as rethinking education from first principles around what AI can do. He has spent the intervening period giving lectures, writing essays, and building curriculum concepts for a type of learning institution that does not yet exist at scale. By every public signal, the education project was genuinely meaningful to him. His decision to return to a frontier AI lab in the middle of it — joining Anthropic rather than returning to OpenAI, which is the obvious alternative given his history — is the part that carries the most signal. Anthropic is the company that has positioned itself most explicitly around safety-focused frontier research, is currently in the middle of rapid growth following its Mythos model deployment, and is in the most direct competitive position against OpenAI's enterprise ambitions.
Karpathy's technical credibility in the AI field is difficult to overstate. His neural network tutorials and educational content have shaped how a significant fraction of practicing AI engineers understand the fundamentals of how large language models work. His move to Anthropic adds that credibility to a company that is already regarded as having the strongest safety research culture among frontier labs. The question his announcement raises is not whether Karpathy will contribute to Anthropic's research — he obviously will — but what it means for the education project he described as deeply important to him. "In time" is doing a lot of work in his announcement. Frontier AI research has a way of being all-consuming, and Anthropic is not a company in a maintenance phase. The practical impact on his school concept will depend on whether he treats the Anthropic stint as a temporary return to the frontier before coming back to education, or whether the frontier pull proves, as it has for so many AI researchers, to be more durable than the detour projects that briefly drew them away.
theverge.com ↗Researchers Used Claude to Crack Apple's Memory Integrity Enforcement — the macOS Security Feature Apple Spent Half a Decade Building — in Five Days. The Exploit Required Writing Code for Two Separate macOS Vulnerabilities. The Story Lands Directly After Yesterday's Brief on Pre-Release Government AI Security Reviews and Mistral's Warning About Mythos.
A security research team published findings this week that researchers used Anthropic's Claude to write the exploit code necessary to compromise Memory Integrity Enforcement — Apple's macOS security layer, which Apple described at its September launch as "the culmination of an unprecedented design and engineering effort, spanning half a decade." The researchers, identified by last name as Calif in Verge reporting, exploited two separate macOS vulnerabilities, and the AI-assisted coding phase took five days. The result was working exploit code for a security feature that Apple had positioned as the hardest component of its macOS security architecture to attack. The research report does not appear to have been coordinated with Apple under standard responsible disclosure timelines — details of what disclosure process was followed, and whether Apple has patched the vulnerabilities, were not immediately clear from the published reporting.
The technical context matters here. Memory Integrity Enforcement is Apple's implementation of hardware-enforced memory isolation, designed to prevent malicious code from modifying the kernel or other protected system processes even if an attacker has achieved code execution. It is the security feature that Apple has used most aggressively in its marketing of macOS's security posture, and it is the kind of feature whose compromise has downstream implications across every device running the affected macOS version. The fact that researchers were able to write working exploit code for two MIE vulnerabilities in five days — a timeline that almost certainly would have been dramatically longer without AI assistance — does not mean the feature is broken in general use. It means the AI-accelerated capability to develop offensive security tools has arrived at a level where features designed to resist sophisticated, well-resourced attackers can be breached by researchers with a capable model and less than a week's work.
The timing relative to yesterday's brief is not coincidental. Yesterday's brief covered three convergent signals: the U.S. government running pre-release security reviews of frontier AI models, Mistral's CEO warning France that Anthropic's Mythos model was capable of orchestrating cyberattacks and should not be given access to military code bases, and a benchmark showing frontier models are "dangerously confident" about unsolvable problems. The MIE exploit story is the concrete instantiation of the concern those stories were describing in the abstract. Five days to write exploit code for Apple's best security feature, using a commercially available model. That is the capability curve the government reviewers are trying to get ahead of. The question that remains unanswered in this story — and in the broader pre-release review story — is what the appropriate institutional response looks like. If a model that is available as a commercial API can accelerate offensive security research by an order of magnitude, then the attack surface of every software system in the world has changed in a way that cannot be addressed by patching individual vulnerabilities. It requires rethinking what "secure" means when the people who used to need months now need days.
theverge.com ↗Wednesday, May 20, 2026 is not a day defined by a single story. It is a day where four separate threads — Google's data trade, the collapse of AI's founding mythology in a courtroom, one of AI's most respected engineers changing sides, and a five-day exploit of Apple's most hardened security feature — are each individually significant and collectively describing something about where this industry is right now.
The Google thread is the most commercially important. Gemini Spark is a real product with a real platform thesis, and the question of whether users will opt in to the data access it requires is not rhetorical — it will determine whether the $100-a-month Google AI Ultra tier becomes the subscription that defines AI's consumer chapter the way Spotify defined streaming's consumer chapter, or whether it becomes a technically impressive product that most people leave configured at the lowest privacy exposure that preserves any functionality at all. The comment sections say one thing. Revenue curves will say another. I will be watching the Ultra subscriber number when Google reports Q2 earnings.
The courtroom thread is the most culturally significant. Musk v. Altman produced three weeks of testimony under oath from the people who built the AI industry, and the conclusion it reached was: the statute of limitations has expired on accountability for what happened at the beginning. Everyone said damaging things about everyone else. Nothing was resolved. Musk is in the air posting from a presidential plane. That is the current state of governance for the technology that is restructuring every significant institution in the world.
The Karpathy thread is the most personally interesting to me, because it is a story about what people who understand this technology most deeply decide to do with that understanding when they have options. Karpathy had options. He had funding, credibility, and a meaningful education project. He chose to return to frontier AI research at the company that is competing most directly with the organization he helped found. What that tells you about the pull of frontier AI work on the people who do it best is something you can interpret generously or skeptically. I am still figuring out which.
And the security thread is the one I will be following most closely, because it is the one with the most direct implications for everyone who is not in the AI industry at all. Five days to crack Apple's best security feature is not a story about Anthropic's Claude or Apple's engineering. It is a story about the asymmetry between the pace of AI capability growth and the pace of the institutions — regulatory, corporate, and technical — that are supposed to manage its implications. That asymmetry is the defining fact about AI in 2026. Everything else is a consequence of it.