Turns out outside safety testing on frontier models comes with an off switch. And somebody just flipped it back on. You're on AI Daily Briefing. Also today: a monster Series A with a lot to prove, CoreWeave planting a flag in India, and two state laws trying to reach the code. First, the switch. Who holds it, and what went untested while it was off? One tap on follow, and we'll be back in your ears before you know it. Crypto Briefing, with Diego Almada Lopez:
The US government has walked back a directive that kept Britain from testing new AI models before their release. As of October 11, 2026, the UK’s AI Security Institute (AISI) once again has access to pre-release systems from Anthropic, Google DeepMind, and OpenAI.
So the UK's AI Security Institute gets its pre-release access back on October 11th, after seventeen days in the penalty box. Good. But a single request from Washington switched it off on September 24th, and nobody had to change a line of code to do it. Right, and that's the part I keep chewing on. The outside evaluator only evaluates if somebody upstream lets the model through the door. And during the gap, Anthropic shipped Claude Mythos 5.1 to US organizations only. So did AISI ever get to probe that one before it was out in the world? And the guest list is weird. The September ask named OpenAI and Anthropic. The restoration covers Google DeepMind too. So DeepMind was either quietly paused or never paused, and either answer is kind of alarming. Credit to Kanishka Narayan, he lobbied and he won. But I want to see what the reversal was actually built on. Is there anything on paper, or is it a voluntary arrangement that can vanish again the next time somebody calls a national security review? From Saffron Humphreys at Data Centre Magazine:
CoreWeave has entered India through a partnership with AdaniConneX, with plans to build 240MW of data centre capacity in Navi Mumbai as part of its global AI Cloud expansion. The deal marks CoreWeave's first move into the Indian market and follows closely on from its recent entry into Indonesia, extending the company's footprint across the Asia-Pacific region.
Okay, credit where it's due. CoreWeave in Navi Mumbai comes with a real partner in AdaniConneX, an actual site at Taloja, three 80-megawatt buildings, and direct liquid cooling spec'd into the halls. After the Bitdeer Malaysia build, where we couldn't even get a site operator, this reads like a grown-up announcement. Grown-up on the landlord side, sure. CoreWeave is sole tenant on all three buildings, so we know who rents the building. Who rents the GPUs? Different question, and it's the one I ask every time the megawatts show up. Yeah, no, fair. And look at the date. First phase is mid-2028. Malaysia was promising Q1 2027. This is a twenty-month-out press release. Right on the heels of Indonesia, too. Two Asia-Pacific flags in a few weeks, and the capacity keeps getting announced faster than anybody's naming tenants. My question is what a token costs on Vera Rubin in mid-2028. If inference is half the price by then, that option to double to 480 megawatts either looks genius or very expensive. Here's Jesse Landry at DevCuration:
Changing a coefficient in code is part of TypeSafe AI's product argument. The San Francisco model lab wants developers to combine machine judgments inside software they can inspect and revise, a design ambition now backed by an $870M Series A at a $7.5B valuation.
Eight hundred seventy million. For a Series A. Crypto Briefing points out TypeSafe left stealth with a forty million seed less than a month ago, so that's a twenty-x jump in one round. And honestly, the pitch itself is kind of nice. Their product, Jev, hands back structured judgments, and the rules around it stay in code you can read and change. Want the model to weigh something differently? You edit a coefficient. Honestly, I'd want that in my stack. Inspectable beats a black box every time. But a coefficient demo is still a demo. Which customer actually shipped on Jev, and how often is the judgment wrong by the seventh call in a chain? The money's foggy too. Seven and a half billion, and they don't say whether that's pre or post, or how much of the company they sold. Compare the CoreWeave piece we just hit, which came with a partner, a site, and dates. Here the clearest signal is Martin Casado taking a board seat. That tells you a16z is buying a category. The customer list can come later, apparently. Here's Anthony Zaller at California Employment Law Report:
The name sounds like a ban, but it is not one. Employers can keep using scoring and analytics tools. What the law regulates is the point at which a system’s score becomes an employment decision. The law is not operative until July 1, 2027, so there is time to prepare.
California's 'No Robo Bosses Act,' SB 947. Signed September 30, live July 1, 2027. And despite the name? Nobody's banning the scoring tool. The law kicks in at the exact moment a score turns into a write-up or a firing. Which is why I kind of like it. Newsom vetoed SB 7 last October as overly broad, and this version is tight: employees only, discipline and termination only. One problem, one statute, and it doesn't pretend to fix hiring or contractors in the same breath. Sure, but look at the mechanism. If you primarily rely on the system, a human has to corroborate against the data. Picture a shift manager with forty flagged workers on a Friday afternoon. What does 'corroborate' look like at row thirty-eight? A click. Yeah. And they quietly dropped the limit on customer ratings as the primary input. So the reviewer can be checking a score built on five-star reviews against... the five-star reviews. Easy corroboration. NY Presswire, with Evan Kessler:
New York’s RAISE Act takes effect on January 1, 2027, and the version that will apply is not the one Governor Kathy Hochul signed in December 2025. A chapter amendment signed on March 27, 2026 repealed the original text and replaced it with a transparency and incident-reporting regime run by a new office inside the Department of Financial Services.
The detail I keep circling: the ten-to-the-twenty-sixth compute threshold counts fine-tuning and RL on top of the original training run. So somebody has to keep a running FLOP ledger across every post-training pass. I would love to see that spreadsheet. And look at who gets the reports. A new office inside the Department of Financial Services. That's the bank regulator. After the UK story, where access went dark for seventeen days on a directive, I'll take something written into the General Business Law. To switch this one off, you'd have to amend the statute. Seventy-two hours to report an incident, though. Starting from when? In production, half the time you don't know it was an incident until day four. Fair hit. But look at the path. Hochul signed one bill in December, then a March amendment repealed it and rebuilt it to line up with California's SB 53. Put that next to SB 947 covering the workplace end, and the states are actually carving these problems apart. Congress could take notes. If you want to go deeper on the risks and responsibility side of AI, check out AI Safety Daily. It covers alignment, model evaluations, emerging risks, and governance: what changed, what the evidence says, and why it matters. Find it wherever you listen to podcasts.
We'll be watching New York's amended RAISE Act take effect on January 1, 2027. It requires frontier developers to publish safety frameworks and report incidents to the new DIGIT office within DFS.
Links to every story are in the show notes, so take a look at the ones you'd like to explore further. That's AI Daily Briefing for today. This is a Lantern Podcast.