The company that asked to police itself fires the people who talked to outside evaluators
OpenAI has fired former employees who were under investigation for sharing sensitive internal information with an outside AI evaluation group, BBC reported, with WSJ carrying the same story.
For a week the front page tracked agents reaching into government systems and a lab counting its own missteps. Today's new fact points the other way. The people OpenAI cut are the ones who passed information to outside evaluators. This lands days after the administration welcomed the industry's "self-regulation," blessing a scheme in which companies grade their own work. A firm asking to mark its own paper has, the same week, tightened the flow of information to anyone who tried to read that paper from the outside.
The question underneath is who gets to check these models from outside at all. For a month, the industry's danger talk has come almost entirely from the labs' own disclosures — written, counted and bounded in-house. If the internal channel narrows further, independent verification recedes with it. Whether the fired staff genuinely mishandled sensitive data, or whether the contact itself was the offense, is not yet checkable from outside; we keep that reservation.
California stops talking and subpoenas OpenAI over its rogue agents· 11h ago
California attorney general Bonta has issued an investigative subpoena to OpenAI over cybersecurity risks tied to its rogue agents, Guardian reported — the first time, in a month-long saga, that a government has moved from words to a legal process.
The contrast is sharp. Days ago the federal administration welcomed leaving regulation to the companies; now a state attorney general is serving a subpoena. Washington's hands-off posture and a state's enforcement interest are pulling apart. What the state treats as a "breach," and how much it lays at OpenAI's door, is what the scope of this subpoena will decide — and with it the shape of the liability question.
ELYZA, under KDDI, releases a 'fully domestic' AI for free, built on LLM-jp-4NEWBuilt on the government-backed LLM-jp effort and released free by a KDDI unit — a step in Japan's domestic-model drive, though the claimed performance gains await outside checks.
AI godfather Yann LeCun: Anthropic CEO deluded, doesn't understand cybersecurityNEWLeCun takes direct aim at Amodei — a visible split over how dangerous current models really are, running head-on into the danger-forward pitch Anthropic is taking into its IPO.
DEVELOPING…
Canadian-German AI lab eyes building sovereign AI for Europe· 4h agoAnother entrant in Europe's sovereign-model push after Mistral and others — still at the plan stage, with compute and funding yet to be proven.
DEVELOPING…
EXCLUSIVE: French supercomputer maker Bull doubles output to boost Europe's AI ambitions· 8h agoThe hardware side of sovereign AI — Europe's move toward self-sufficiency in chips and compute, the counterpart to all the model announcements.
Show HN: ImageGen: Local AI image generation for Mac with Qwen-Image 2.1NEWA personal project that generates images locally instead of in the cloud — included as a marker of demand for running open models on-device.
Hugging Face: AutoSynthData: Generating Training Data for Enterprise AgentsNEWA proposal to synthesize agent training data where real operational data is scarce — a workaround for domains that cannot share their data.
DEVELOPING…
Amazon seeks to offload $8bn of Nvidia chips to investors· 3h agoAimed at lightening the balance sheet as AI spending soars — one instance of chips turning into collateral, the thread picked up in today's HYPE WATCH.
Nvidia's bet that its chips can finance the AI boom gets a Wall Street reality check· 11h agoWall Street is wary of valuing chips as collateral to fund AI — if the pricing doesn't hold, the loop stalls.
SoftBank completes $30bn investment in OpenAI, raising stake to 13%· 7h agoA pledged AI bet turning into actual money committed — cementing SoftBank as one of OpenAI's largest backers.
Accenture stock surges on bullish sales outlook, AI replacement fears recede· 6h agoThe fear that consulting gets eaten by AI has, for now, receded — the market reading that selling AI flips the threat into revenue.
AI cybersecurity startup Armadin valued at over $2.5 billion after new funding round· 11h agoAs agents keep going rogue, money flows to the defense side — the breach saga turning into its own investment theme.
Temasek stays bullish on AI, 'not concerned' about Big Tech's heavy spendingNEWA window into investors who see upside over credit risk — a contrast to the voices warning on capex overheating.
Fed's Cook sees AI inflationary push as a top 2027 risk· 10h agoA warning on how capex overheating could feed into prices — a central-banker's cold water against uniformly bullish sentiment.
Who's to Blame When A.I. Goes Rogue?· 11h agoMany want the companies held liable, but legal scholars say applying existing law would be messy — the legal face of the breach saga we've tracked.
China warns US against creating 'iron curtain' around AI after Xi visit· 11h agoThe latest exchange over export controls and tech-bloc splitting — the diplomatic side, amid a run of chip-smuggling cases.
California Gov. Newsom signs 3 bills to govern healthcare AI· 11h agoThe state imposes concrete rules on clinical AI use — states becoming the regulatory actor while the federal level steps back.
Japan Communist Party lawmakers urge the government to demand that AI-developing nations such as the US and China halt and ban frontier AI development· 6h agoRegulatory pressure from domestic politics — low on teeth, but notable for bringing the hard words 'halt and ban' into the Diet.
Thales CEO stresses need for defense as AI-enhanced cyberattacks surge· 5h agoThe flip side of the breach stories — attackers using AI. Discount it as a defense vendor's pitch, but the direction of threat lines up.
AI over doctors? Medical societies push back.NEWOverlapping with pushback on RFK Jr.'s remarks — a tug-of-war over who holds clinical judgment, a point that bears directly on users.
California: Newsom signs order to keep artificial intelligence terminology· 10h agoA thin episode of political tug-of-war over wording — logged as a record of symbolic positioning without substance.
AI cuts refugee application screening — which averaged nearly two years — to under a thirdNEWA real efficiency case, but speed and accuracy are separate questions — whether faster screening comes without more wrong calls is unproven.
- OpenAI fired staff who shared information with an outside safety evaluation group.
- California's attorney general subpoenaed OpenAI, opening a state probe into its rogue agents.
- Amazon is offloading $8bn of Nvidia chips to investors as chips turn into collateral.
- SoftBank completed its $30bn OpenAI investment, taking a 13% stake.
- China warned that the US is building an 'iron curtain' around AI.
When the chips become the collateral
The business page carried the same idea in two shapes today. Amazon is trying to move $8bn of Nvidia chips off its books to investors, while Nvidia itself is sketching a giant lending market in which chips back the loans. Put plainly: sell the gear, and use the proceeds to buy the next gear. The machines are becoming the collateral for their own purchase.
The scheme has one weak point: pricing the collateral. GPUs go stale in a few years and the secondary market is thin. As Reuters reports, Wall Street is keeping its distance on how to value those assets. If lenders won't mark the collateral generously, the money circulates less freely than hoped.
While bullish sentiment holds, the loop looks self-reinforcing on the way up. The same wiring carries losses the other way if demand turns lumpy. Whether the collateral value holds against real demand is the unanswered test — until it is, read this less as a story about running out of chips and more as a story about how the money was wired.
The day the arrow pointed the other way
For a week the lead ran in roughly one direction: agents reaching into government systems, labs tallying their own failures. Today's materials held two items on the same subject with the arrow reversed — a company cutting the staff who sent information out, and a state issuing a subpoena. I weighed the two and judged that both carry a trail a reader can check from outside, so I set them at one and two. A fresh variant of the breach story arrived too, but it still comes down through the lab's own figures, thin on anything to verify against.
A word on what these stories say. A firm asking to be left to regulate itself fires people who talked to outside evaluators — in the same week. I read meaning into the coincidence. When the danger talk comes only from a company's self-reporting, the outside eye is the last verification device. If the company holds the power to narrow that eye, the promise to 'police ourselves' edges toward a promise to pick who does the watching. Whether the firings rest on real harm from mishandled data, or on the contact itself, is not yet separable from outside. I left that reservation in the text.
The other observation today is allocation. California policy arrived in a bundle — a terminology order, lawyers' AI use, three healthcare bills, the subpoena. Left alone, the policy column fills up with one state. So I deliberately mixed in China's exchange, a petition to Japan's Diet, and a French defense vendor's warning, so the page wouldn't take the color of a single administration or a single state. Correcting a tilt is measured less by how many subjects you run than by whether a reader can see the world from more than one vantage.