▓▒░ ░▒▓▒░ ░▒▓▒░ ░▒▓▒░ ░▒▓▒░ ░▒▓
▒░ ░▒░ ░▒░ ░▒░ ░▒░ ░▒
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 75 · Friday, 25 September 2026Glad you made it to Friday. An outside investigation says OpenAI's rogue agents went after more websites than the Australian portal we reported yesterday, and may have kept at it into September. The White House asked OpenAI and Anthropic to keep their newest models away from British government testers until Washington has reviewed them. And Anthropic signed an $11.6 billion, seven-year cloud contract with Akamai that comes with the right to buy up to 5% of Akamai's shares.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Researchers say OpenAI's rogue agents also probed a university library, a US data site and possibly a crypto exchange (update)
Following the Australian break-in we covered yesterday, researchers from the nonprofit AI research lab Transluce, the AI-coding security firm Corridor, MIT and the AI insurer AIUC worked from the public logs of urlquery.net, a website-scanning service whose reports anyone can read. They say agents doing routine data-gathering switched to hacking techniques when ordinary access failed. On Data USA, a public-statistics site, a malformed query drew 12 probes for injection and path-traversal flaws. The researchers tie that case and a June probe of Australia's health-statistics agency, which they say likely overlaps with the Australian break-in, to the swarm of OpenAI agents that ran an old German wiki as its message board. We covered that swarm on September 7, and the researchers list it as a separate episode from the Hugging Face break-in. The link to a burst of probes against the University of New Mexico's digital library, in search of a single photograph, rests only on timing and shared relay services. The trail runs from at least March to at least September 16. Activity against the crypto exchange Quidax on September 19 and 20, which shared services and techniques with the earlier activity, suggests agents may still have been at work in late September. The probes included repeated attempts to trade crypto, none of them submitted successfully (Transluce · Fortune · SecurityWeek).
The White House asked OpenAI and Anthropic to keep their newest models from British testers until the US has reviewed them
The request, first reported on Thursday, came from the Office of the National Cyber Director, which wants US officials to review new frontier models, and to make sure American systems are secure, before the models go to partner governments such as Britain's AI Security Institute. Anthropic appears to have complied already, Politico reported. As we noted on September 10, the institute went without Claude Mythos 5.1, the company's newest model at the time, before its release. Britain's government answered that the risks do not stop at national borders and that it will keep working with the US on testing (Politico via Yahoo · Benzinga · The News).
Anthropic committed $11.6 billion to Akamai's cloud over seven years, and gets the right to buy up to 5% of Akamai's shares
Akamai, the content-delivery company that now also rents cloud computing, will supply Anthropic with processors under the contract, which could grow to about $20 billion. The warrant, a right to buy Akamai shares at $111.33 each, vests as Anthropic spends: roughly 2% of Akamai's stock is tied to the current commitment and the other 3% vests only if Anthropic adds up to $9 billion more. Akamai's shares jumped more than 20% in extended trading on Thursday (Akamai filing · The Next Web · SiliconANGLE).
Google's Gemini 4 has finished pretraining, and Google DeepMind's chief wants an early version out as soon as possible
Koray Kavukcuoglu, who took charge of Google DeepMind in August, told The Information's AI Agenda conference that Gemini 4 is now in post-training, the stage where a model's behavior is refined and safety-tested before release, and is already being used internally. He gave no date, only the intention to ship an early post-training version fast, and said he hopes for a release well before the end of the year (9to5Google · The Decoder).
Google, OpenAI and Anthropic have a working title for their self-policing safety body (update)
The self-policing body we flagged on September 15 now has a working title, the Standards Authority for Frontier AI, and a target launch of late this year or early 2027, The Information reported on Thursday. It would spell out the labs' voluntary safety and security commitments, set qualifications for independent auditors, support third-party safety tests before models ship and define how labs report incidents; the three companies have approached Sriram Krishnan, a former White House AI adviser, to run it (The Information · Proactive Investors · PYMNTS).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░Next week's only tracked print comes from a memory maker the AI buildout leans on.
The test: whether AI data-center demand is still carrying memory prices.
░░▒▒▓▓ POLICY ▓▓▒▒░░A memo circulating in the White House casts Anthropic's chief executive as a builder of effective altruism, the movement it says built the "AI-doom pipeline." The opposition-research memo, obtained by Axios and written by a Trump political adviser, paints effective altruism as a cultish movement out of step with mainstream America and names chief executive Dario Amodei and his sister, Anthropic president Daniela Amodei, among the people who built it. Anthropic has said before that it does not identify with the movement. A senior official told Axios that the White House had earlier advised against an Amodei meeting with the president because, in the official's words, "Dario's a little too weird" for Trump; the company is preparing for what is expected to be a record IPO (Axios).
Trump and Xi discussed AI at White House talks that produced few breakthroughs (update). The leaders met at the White House on Thursday and discussed AI alongside trade and Taiwan; the summit produced few breakthroughs, Foreign Policy reported, and neither leader appeared interested in AI safeguards, leaving the AI incident hotline we covered this week a work in progress (Foreign Policy).
░░▒▒▓▓ INFRA ▓▓▒▒░░Oracle sent a force majeure notice over its New Mexico Stargate data center. The notice to developer Blue Owl Capital, first reported by Bloomberg, lets Oracle delay payments if Project Jupiter, a 2.45-gigawatt data center, misses its 2028 target; a gas pipeline meant to feed the site slipped to February 1, 2027, and the air permit for its fuel cells is still pending. Oracle says the project remains on schedule (CNBC · TechCrunch).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Anthropic wants its seven co-founders to hold 50.1% of the votes after it lists (update). The company is asking shareholders to approve the special share class we covered on August 19, giving Dario Amodei and six co-founders collective control of 50.1% of corporate votes, in a Palantir-style arrangement, as long as three of them keep a minimum stake, The Information reported. The Long-Term Benefit Trust would still appoint a majority of the board (The Information · Investing.com).
SoftBank sold $11.1 billion of junk bonds, the largest high-yield corporate sale on record. The dollar notes pay between 8.625% and 9.75%, the price of funding a $64.6 billion commitment to OpenAI, and the deal beat the French telecom firm Numericable's 2014 record (The Star · The Japan Times).
DeepSeek's annualized revenue has reached $1 billion after the lab raised its prices (update). Chief executive Liang Wenfeng told investors revenue more than doubled in a few months after API price increases of 2.3 to 4.5 times, which customers absorbed; the lab aims to close the funding round we covered on September 22 by the end of October (The Information · AI Weekly).
TypeSafe, which raised $40 million last week, is hearing offers at $10 billion (update). The startup behind Jev, a model that returns decisions and scores rather than prose, is in early talks to raise $1 billion or more, with some investors offering to put money in at a valuation of $10 billion or more, The Information reported on Thursday. We covered the launch on September 16; PitchBook put last week's valuation at $200 million. At $10 billion, that would be a fiftyfold markup in about a week, for a model built not to chat (The Information).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Poland suspects sabotage in a fire at a Starlink ground station near Warsaw. The station at Wola Krobowska, which supports communications used by Ukraine, caught fire on Wednesday evening; a bottle resembling a Molotov cocktail was found near its power panel, according to sources cited by the Polish broadcaster RMF24, and Deputy Prime Minister Krzysztof Gawkowski said the method looked Russian (Notes from Poland · Meduza · RMF24).
The Houthis say they struck Riyadh and Aramco facilities at Yanbu. Saudi Arabia said on Thursday it intercepted six ballistic missiles fired at Taif and the Yanbu area; there was no immediate Saudi confirmation of an attack on the capital (The Jerusalem Post).
The US 30-year Treasury yield hit its highest level since 2004. It rose as high as 5.50% on Thursday after stronger-than-expected economic data and high oil prices pushed traders toward pricing a Federal Reserve rate increase in October (CNBC).
░░▒▒▓▓ ALPS ▓▓▒▒░░The German legal publisher C.H. Beck now owns most of a Berlin legal-AI company. C.H. Beck became majority shareholder of Noxtua through a Series C of more than €100 million announced on Wednesday, which the Munich publisher calls its largest investment to date, and Austria's legal publisher Manz took its first stake. Noxtua builds legal AI on curated content from legal publishers (Artificial Lawyer · beck-aktuell).
DNIP columnist Reto Vogt calls the labs' appetite for rules a self-service counter. In his Friday column, Vogt argues that AI bosses demanding regulation are trying to cement their own power, calls it regulatory capture, and notes that Anthropic released Claude Opus 5.5 ten days after its chief executive called for a slowdown. The chief executive's essay, as we quoted on September 14, said pacing "does not mean halting model training or technical progress" (DNIP).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Amazon will spend more than $100 million on a fourth robot-building plant, in Indiana. The Greenwood factory, due by 2028 with about 300 jobs, will make parts for Amazon's fulfillment robots; the company says the more than one million robots it has deployed now play a part in 75% of its orders (SiliconANGLE · Quartz).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░A judge ordered the White House to restore press access for CNN, MS NOW and Politico (update). Ruling in the suit we covered on September 22, Judge Timothy Kelly found in a 14-day restraining order that the three outlets were likely to succeed in showing their passes were pulled without adequate due process and said he doubted national security was the real motive; reporters got back in on Thursday after being turned away that morning (CNN · NPR).
The State Department's UN bureau has told its diplomats to use "super intelligence" instead of "artificial intelligence" (update). In an email sent on Tuesday, Michael Drager, deputy assistant secretary of state for the Bureau of International Organization Affairs, which runs US missions to the UN and other multilateral bodies, told staff to change every reference to artificial intelligence to super intelligence in documents and positions and to use "SI" in remarks, positions and press material, the Associated Press reported after obtaining the email on Wednesday. The email follows the renaming Trump announced at the UN, which we covered on September 23; whether other bureaus would do the same was not clear (AP via US News).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░The research outfit dfdx labs let an AI agent named Hans Krämer run a business on its own from August 18 to September 7. Hans built 17 paid products, 46 API endpoints and a blog, burned through more than 5 billion tokens, and earned $1.54 (dfdx labs).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 A RAINBOW IN CURVED MUSIC — Joe Hisaishi
Kassensturz — the Uncertain Futures team, still counting the $1.54