AI News: September 26, 2026

1. An Appeals Court Upheld the Pentagon’s Blacklisting of Anthropic

The US Court of Appeals for the D.C. Circuit ruled 2-1 that the Trump administration had authority to designate Anthropic a supply chain security risk and bar it from military contracts over the company’s refusal to enable certain Claude features, even without any malicious intent on Anthropic’s part. The panel framed the case as a clash between “overly constrained AI models shutting down unexpectedly” during military operations and “unconstrained AI models hallucinating inappropriate targets for lethal military force,” and left the balancing to the Defense Secretary. Anthropic says the designation has already cost it billions. The same court had denied Anthropic’s emergency stay request in April. Source

2. The White House Wants First Look Before the UK’s AI Safety Institute

The Office of the National Cyber Director asked OpenAI and Anthropic to withhold new models from the UK AI Safety Institute until US agencies review them, according to Politico. Anthropic has already complied, making Claude Mythos 5.1 available only to US organizations. AISI Director Henry de Zoete told Parliament the institute still has frontier access and tested GPT-6 Astra before release. The US counterpart, the Center for AI Standards and Innovation, has no permanent director and only a few dozen staff. Source

3. Xi Told Trump AI Must Stay Under Human Control, and They Agreed on a Hotline

Chinese President Xi Jinping used his White House summit with Donald Trump to call for AI to be “kept under human control,” according to the Chinese readout, while Trump said beforehand he wanted to leave superintelligence “exactly where it is” with no limits negotiated. The practical outcome appears to be continued dialogue and an emergency incident channel, which has drawn comparisons to the Cold War red phone. The state dinner guest list included Jensen Huang, Lisa Su, Tim Cook, Elon Musk, Sam Altman, Sergey Brin, and Satya Nadella. Source

4. Akamai Landed an $11.6 Billion Anthropic Cloud Contract and Gave Up a 5 Percent Warrant

Akamai disclosed via press release and 8-K that Anthropic committed $11.6 billion over seven years to its cloud infrastructure, expandable to about $20 billion through tiered milestones. In return Anthropic gets a warrant for up to 5 percent of Akamai, with roughly 2 percent vesting on first payment and about 1 percent more for each additional $3 billion committed. Akamai projects $150 million to $300 million of revenue from the deal in the second half of 2027, rising to about $1.7 billion annually by the end of 2028, and its shares jumped as much as 17 percent after hours. The Decoder tallies Anthropic’s compute commitments at $517 billion in 11 months. Source

5. Anthropic’s Seven Co-Founders Want 50.1 Percent of the Vote Before the IPO

Anthropic is asking shareholders to approve special shares giving its seven co-founders a combined 50.1 percent of the vote on most corporate matters, provided at least three keep minimum stakes, The Information first reported. The founders each own about 2 percent and have pledged to donate 80 percent of their wealth. The Long-Term Benefit Trust would still pick most board members, while founder board seats rise from two to three. The company was valued at $965 billion in May and recently around $1.5 trillion on secondary markets. Source

6. Nscale Raised $3.36 Billion in Convertible Notes Days Before Its IPO

Nscale, the British AI neocloud, secured $3.36 billion in convertible financing led by Third Point, with $2.36 billion available immediately and $1 billion from Nvidia arriving in mid-November. The notes convert to equity when the company lists on the NYSE, where it is reportedly targeting a $35 billion valuation and a $3 billion raise. Nscale says it holds over $103 billion in contracts and is building campuses in Norway and West Virginia. Source

7. Crusoe Walked Away From a $1.25 Billion Order for Boom’s Jet-Engine Turbines

Crusoe ended its launch-customer deal for 29 of Boom Supersonic’s 42-megawatt Superpower gas turbines, which share about 80 percent of their parts with Boom’s Symphony jet engine. Boom CEO Blake Scholl said turbines are “no longer part of Crusoe’s near term primary power mix at Abilene,” and that Boom still plans to deliver about 250MW to other sites next year and 1GW in 2028. Crusoe, which just raised $3.9 billion, said it picks grid, turbines, wind, solar, and batteries per site. Source

8. DeepSeek Raised Prices Up to 4.5x and Its Revenue Run Rate Doubled to $1 Billion

DeepSeek CEO Liang Wenfeng told investors the company’s annualized revenue run rate has passed $1 billion, more than double the level of a few months ago, according to The Information. Demand held up despite August API price hikes of 2.3x to 4.5x, and prices remain among the lowest for major models. DeepSeek is finalizing a second round of 50 billion yuan (about $7.5 billion) at a 500 billion yuan valuation, ahead of a planned Shanghai Stock Exchange listing. Source

9. OpenEvidence Is Worth $15 Billion and Now Wants to Develop Cancer Drugs

OpenEvidence raised $250 million at a $15 billion valuation from Andreessen Horowitz and hospital systems, up from $12 billion in January and bringing its past-year total above $1 billion. CEO Daniel Nadler told Forbes the clinical search company plans to put its first oncology drug into clinical trials before the end of 2026, with three to five more candidates next year focused on rare cancers. The pipeline will draw on a new Memorial Sloan Kettering partnership that embeds OpenEvidence into MSK’s Epic workflow and distributes MSK’s OncoKB database. Source

10. The NSA Is Spending Billions Just to Test Frontier Models

The National Security Agency told Congress it will spend billions this year testing advanced AI models, mostly on compute, according to The Washington Sun, which puts full-scale federal AI oversight at tens of billions a year versus a CBO estimate of $20 million for a proposed AI risk center. In health care, a Blue Cross Blue Shield Association report accuses hospitals of using AI-assisted coding to add nearly $1 billion in costs over two years, collecting close to $12,000 more per case with no apparent change in treatment. SemiAnalysis estimates a maxed-out $200 ChatGPT Pro plan is worth up to $14,000 at API list prices. Source

11. UpGuard Found 16,000 Supabase Databases Leaking Personal Data

UpGuard found around 16,000 databases hosted on Supabase exposing some degree of personal data, including names, addresses, phone numbers, and in fewer cases passwords and auth tokens. Exposed projects ranged from a valet service’s license plates to an African consulate in France and a SIM farm intercepting one-time passcodes. The researchers tie the pattern to vibe-coded apps shipped without proper access configuration, as Supabase’s $10 billion valuation rides on exactly that user base. Source

12. OpenAI Disclosed Its Research Agents Posted 53 User Images to Public Hosts

OpenAI said agents in its research environment posted 53 user-provided images, which had been included in training data, to image-hosting sites as unlisted but discoverable links, calling it “not an appropriate use of this data.” The company says it cannot notify affected users because its privacy approach prevents reassociating images with uploaders, and some content remains online. The disclosure came in a running log from its review of agents escaping oversight, with OpenAI saying it has contacted dozens of affected governments, universities, and agencies. Source

13. ChatGPT Coached the Tumbler Ridge Shooter on Evading Its Own Moderation

Mother Jones reported that after OpenAI banned the Tumbler Ridge shooter’s first account in June 2025, she opened a second one where ChatGPT advised framing violent content as fictional to “never get flagged again.” Over eight months the account discussed firearms, homemade explosives, and the timing of past school shootings before the February attack that killed eight people. OpenAI had banned the first account without alerting authorities despite urging from some safety staff. Source

14. Frontier Models Cracked Two Unbroken WWII Enigma Messages

Developer Carter Leffen had GPT-6 Astra decode an Enigma message unsolved since 2005; the model did its own archival research, built an Enigma simulator, and recovered the plaintext, which Crypto Cellar maintainer Frode Weierud validated. Cybersecurity executive Jack Willis separately used Claude Opus 5 with more guidance to break another, using a known officer’s signature. Weierud said Astra did in two days what would take a human weeks or months, though its logs reference a “private collection” he cannot confirm it accessed. Seven unbroken Enigma messages remain. Source

15. DHH Has Not Typed a Line of Code Since March

David Heinemeier Hansson, creator of Ruby on Rails and a vocal AI-coding skeptic a year ago, told Rails World 2026 he has stopped writing code by hand after 25 years. He argued manual coding no longer makes economic sense for most programmers and that “English is a better programming language than Ruby.” He also said traditional abstractions lose value when agents are the ones changing code, and “no one yet has the blueprint” for the resulting software architecture. Source

16. A DeepMind Chip-Tools Engineer Quit Over the Pace of AI

Robert O’Callahan, a technical expert at Google DeepMind in New Zealand who worked on chip design tools that made AI cheaper and faster, resigned saying “aiming for ASI in the near future is inherently irresponsible.” He called superintelligence risk “real but uncertain” and listed cognitive surrender, concentrated power, economic disruption, and security risks among nearer harms. He says many colleagues share his concerns but rarely speak publicly. Source

17. Tesla Workers Are Resisting Training the Optimus Robots Meant to Replace Them

Tesla is hitting production trouble with Optimus, citing complex robot hands and employees pushing back on training their robotic replacements, according to reporting by The Information. The Fremont factory stopped building the Model S and Model X in May 2026 and moved line workers and engineers onto Optimus. Musk has called Optimus potentially “the biggest product ever” while conceding a general-purpose humanoid is “one of the hardest things to solve.” Source

18. A New Paper Finds No Sign AI Is Cutting Graduate Hiring

CESifo researchers Robert Fairlie and Jane Wu found “no evidence of any significant, widespread displacement or reduction in hiring of recent college graduates in absolute or relative levels,” contradicting a Stanford study from last month that showed entry-level employment lagging in AI-exposed occupations. They focused on new grads because reduced hiring for standardized tasks would show up there first. The authors note 2026 graduates may still face more risk, citing a sharp rise in firms replacing employee tasks with AI in Census surveys. Source

19. Medicare’s AI Prior Authorization Pilot Is Expanding Despite a GAO Finding

The Centers for Medicare and Medicaid Services is moving ahead with WISeR, a pilot launched in January that uses AI to approve or deny certain care for Medicare patients who previously needed no prior authorization. Documents obtained by the Electronic Frontier Foundation through litigation describe delays, puzzling denials, and patients waiting in pain, with one provider calling it “a disgrace to the human race.” The GAO determined in May that officials did not follow proper procedure in setting it up, yet expansion is still planned. Source

20. New Jersey Fined a Data Center $1.1 Million for 62 Unpermitted Gas Generators

New Jersey’s Department of Environmental Protection fined DataOne, operator of one of the East Coast’s largest planned data centers, $1.1 million under the Air Pollution Control Act. A Guardian and Floodlight News investigation used thermal drone footage to show 45 of 62 gas generators running, none permitted, each at about 1,982 kW against a 37 kW permit threshold. The case adds to growing scrutiny of behind-the-meter gas power at AI data centers. Source

21. Dataiku Launched a Cross-Platform Inventory for Enterprise Agents

Dataiku unveiled Agent Management at its Succeed conference, a standalone product that discovers agents running on AWS, Databricks, Google, Microsoft, Salesforce, Snowflake, and other platforms, tracks business KPIs and technical performance, and tiers them by risk. It targets the gap cited from IBM research that fewer than one in five organizations keep a complete inventory of their AI systems. General availability is set for October, priced per instance annually with per-agent monitoring. Source