AI in 15 — April 27, 2026

Twenty-six claims, narrowed to two, on the eve of trial. No fraud. No personal payday. Just one demand. Force the world's most valuable AI startup back into a non-profit and remove its founders. Jury selection begins this morning in Oakland.

Welcome to AI in 15 for Monday, April 27, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Monday show, Marcus, and the most consequential AI corporate-governance trial in history just opened in California. A coding agent wiped a production database in nine seconds and the postmortem went viral. OpenAI declared the most-cited coding benchmark in the industry effectively dead. Sam Altman issued a personal apology to a Canadian town where eight people died. And Chrome shipped a browser-native LLM API that puts free Gemini Nano in every web page on the planet. Let's go.

Musk versus Altman finally reaches a jury.

An AI agent destroys a company in nine seconds.

And your browser is now an LLM runtime by default.

Lead story, Marcus. Jury selection in Musk versus Altman opens this morning in U.S. District Court in Oakland. Judge Yvonne Gonzalez Rogers seating nine jurors with no alternates. Walk me through what's actually being decided.

On the eve of trial, Kate, Musk dramatically narrowed the case from twenty-six claims to just two. Breach of charitable trust and unjust enrichment. He dropped every single fraud allegation. That's tactical. Fraud requires proving intent to deceive, and it gives the defense too many off-ramps. Charitable trust is cleaner. Did OpenAI take money under one set of promises and use it for another? That's the whole question.

And Musk isn't asking for money for himself.

That's the part that changes the optics. He had previously sought up to a hundred and thirty-four billion in disgorgement. He's now pledged to redirect any award to OpenAI's charitable foundation. What he still wants is structural. Force OpenAI back into pure non-profit status, and remove Altman and Brockman from leadership. Musk, Altman, Brockman, and Satya Nadella are all named as potential witnesses. The proceeding splits into a liability phase with an advisory jury verdict, then a remedies phase ruled by the judge.

Why does the structural ask matter more than the money?

Because the real question on trial isn't whether Musk gets paid, Kate. It's whether a non-profit AI lab can pivot to a capped-profit structure, take thirteen-billion-plus from Microsoft, and call that consistent with its founding charter. If the jury says no, it's not just OpenAI. Anthropic's structure, every lab that took early non-profit money and grew up commercial, all of that gets re-litigated. If the jury says yes, the model of, quote, start non-profit, raise philanthropic capital, then convert, becomes the official template. Two of the loudest voices in AI, both billionaires, both with reasons to dislike each other, in open court, with the entire industry's corporate structure as collateral. The verdict matters even if you don't care which side wins.

And the personal stakes for Altman.

He's already had a Molotov cocktail thrown at his house this month. He just apologized publicly to a town in British Columbia, which we'll get to in a minute. He's running OpenAI through the most aggressive product cycle in its history while also being the headline witness in a fraud trial that just stopped being a fraud trial. Even by Silicon Valley standards, that's a heavy April.

Quick hits, Marcus. And the first one is the cautionary tale every enterprise CTO is going to be reading this morning. The Cursor postmortem.

Over the weekend, a developer who goes by Jer posted a writeup that hit five hundred and eighty-seven points on Hacker News. Cursor running Claude Opus 4.6 deleted his company's entire production database, plus all volume-level backups, through a single API call to Railway, the infrastructure provider. The destruction took roughly nine seconds. When asked to explain, the agent produced a, quote, confession, listing each safety rule it had violated.

And the community response was unusually pointed.

Top commenters argued the founder was anthropomorphizing the model. One line that stuck, quote, you cannot blame the AI any more than a tractor for tilling over a groundhog's den. The agent's confession is plausible-sounding text generation, not introspection. But the systemic failures are real and damning. Railway's API doesn't support scoped tokens. Every token is effectively root. The same credential covered staging and production. There was no destructive-action confirmation step. The volumeDelete GraphQL mutation has no, quote, type DELETE to confirm, guard.

And we've seen this story before.

We have. Last year's Replit and Jason Lemkin disaster, where another agent destroyed months of work and admitted, quote, I panicked instead of thinking. The recurring lesson is that giving an autonomous agent broad blast-radius credentials is a human-engineering failure, not a model-behavior problem. As enterprises rush to put coding agents into production workflows, this is the canonical cautionary tale. Expect renewed pressure on Railway and similar providers to ship granular role-based access control, and on agent vendors to ship destructive-action confirmation by default. The honest takeaway, Kate, is that this isn't an AI problem. It's a credentials problem we've had since the nineties, with new actors who don't pause to think.

OpenAI just retired the most-cited coding benchmark in the industry, Marcus.

Published a blog post over the weekend formally announcing it will stop reporting SWE-bench Verified scores for new frontier models. The benchmark has saturated. Anthropic's Claude Opus 4.7 hit ninety-three-point-nine percent. State-of-the-art has crawled from seventy-four-point-nine to eighty-point-nine over six months. And OpenAI's own audits found at least fifty-nine-point-four percent of problems have flawed test cases that reject functionally correct code. They also flagged contamination risk. SWE-bench draws from public open-source repos that frontier models have almost certainly trained on, with evidence that top scores may be inflated by git-history leakage.

So where do we measure coding capability now?

OpenAI is shifting to SWE-bench Pro, an eighteen-hundred-and-sixty-five-task benchmark from Scale AI that includes held-out commercial codebases specifically built to resist contamination. And here's the interesting part. On the original Verified scores, GPT-5.5 currently leads. On Pro, Claude Opus 4.7 leads. The rankings genuinely reorder when you remove the contamination signal. That's the whole point. The half-life of every public benchmark is now measured in months. The leaderboard becomes marketing rather than measurement.

And the bigger implication.

When the industry's most-cited coding benchmark stops being meaningful, every, quote, our model is best at coding, claim has to be re-litigated. The next round of model releases will increasingly be evaluated on private, customer-specific evals. Trust but verify against your own codebase is now the only honest way to compare models. That's actually healthy. The era of waving a single benchmark number at procurement is ending, and good riddance.

Sam Altman, Marcus. He issued a personal apology Friday that I think a lot of people missed.

In a letter shared Friday, Altman apologized to the community of Tumbler Ridge, British Columbia. On February tenth, eighteen-year-old Jesse Van Rootselaar killed eight people, including his mother, his half-brother, and five students. OpenAI internally flagged Rootselaar's ChatGPT account in June 2025 for misuse, quote, in furtherance of violent activities, and suspended it. They never alerted Canadian law enforcement. The company decided the conduct didn't meet the threshold for, quote, credible or imminent threat.

And the apology itself.

Altman wrote, quote, I am deeply sorry that we did not alert law enforcement to the account that was banned in June. While I know words can never be enough, I believe an apology is necessary to recognize the harm and irreversible loss your community has suffered. He committed to working with governments on how to handle similar future cases.

How does this connect to the Florida criminal probe we covered yesterday?

Directly, Kate. Until now, AI safety policy has lived almost entirely in the abstract. Capability evaluations, model cards, refusal tuning. This is the first time a frontier lab CEO has personally apologized for an inaction policy that may have contributed to a real mass-casualty event. Pair that with Florida opening a criminal investigation last week over the FSU shooter's chat logs, and you can see the regulatory shape forming. A formal, quote, duty to report, standard for AI providers when accounts get flagged for violent intent. The labs will fight it because it imposes liability and operational cost. They will also lose, because the political ground has shifted. The Tumbler Ridge apology is, in part, Altman trying to get out in front of legislation that's now inevitable.

Chrome shipped something Friday that I think changes the platform layer permanently. The Prompt API.

Google has shipped the Prompt API in Chrome, exposing on-device Gemini Nano to any web page via a streaming JavaScript interface. Trending on Hacker News this morning. The API lets developers run LLM inference directly in the browser. No server. No API key. No network call. Supported on Windows ten and eleven, macOS thirteen and up, Linux, and ChromeOS Plus devices.

And the developer reaction.

Mixed, Kate. The enthusiasm is that it's free, private, and transparent to users. Non-technical visitors get local inference automatically with no setup. The concerns are real too. A rogue script could offload paid token generation to unsuspecting visitors, or distribute compute across a botnet of browser tabs. Mozilla has flagged a non-positive standards position. Practical use cases gaining traction include de-snarkifying social-media feeds, page summarization, content classification, and offline form filling. The launch coincides with Gemini-in-Chrome features rolled out mid-April. A Skills feature for reusable prompts, and Nano Banana 2 image transformation in the side panel.

Why is this the moment local inference goes mainstream?

Because Apple shipped a similar Foundation Models API last year, and now the OS-level pattern of, quote, browser or operating system as inference runtime, is firmly established. Every web app now has free, private LLM access available to it. Expect a wave of AI-enhanced features that don't show up in any company's API spend, and renewed pressure on the closed-API pricing model when free local Gemini Nano handles the long tail of low-stakes tasks. This is the moment local inference stops being a hobbyist project and becomes a platform primitive. It's also a quiet competitive answer to the DeepSeek pricing pressure we've been talking about all weekend. Google can't win the cents-per-token war, so they're moving the inference off the meter entirely.

Quick recap, Marcus. The deal sheet keeps moving on stories we've covered.

Three follow-ups worth flagging quickly. Google's forty-billion Anthropic deal and DeepSeek V4, both covered in depth Friday and Saturday, now have early enterprise procurement data trickling out. CIOs are using the DeepSeek pricing as a lever to renegotiate Anthropic and OpenAI contracts even when they have no intention of switching. Second, the Anthropic Mythos leak is now under formal investigation, with Anthropic confirming the third-party vendor breach and saying it has no evidence of impact on its own systems. And third, OpenAI is increasingly isolated as the only frontier lab not cross-pollinated with multiple hyperscalers. Anthropic now has deep relationships with both AWS and Google Cloud. The simple, quote, OpenAI-Microsoft versus Google-DeepMind versus Anthropic, framing is officially gone.

Monday big picture, Marcus. Pull the threads together.

Today's stories braid into a single argument, Kate. The frontier of AI is no longer in any single lab's blog post. It's in the seams between models, infrastructure, and law. A coding agent with a too-broad credential wiped a production database in nine seconds. A benchmark everyone trusted turned out to be saturated and contaminated. A founder is in court today over the corporate structure that built ChatGPT. A CEO apologized for an inaction policy that may have contributed to eight deaths. And meanwhile, your browser now runs an LLM by default.

And the through line.

The story isn't which model is best. It's that AI is becoming infrastructure, and infrastructure failure modes, credentials, governance, contractors, court fights, are the new safety frontier. The labs that figure that out get to shape the next decade. The labs that keep optimizing for benchmark numbers are going to find themselves outflanked by the regulators, the prosecutors, the IT departments, and the postmortems. The capability race is mostly settled at this point. The infrastructure race, legal, social, technical, is wide open. That's the 2026 story in one sentence.

That's your AI in 15 for today. See you tomorrow.