AI News: September 28, 2026

1. Amodei Had His First One-on-One Dinner With Trump Ahead of a White House AI Announcement

Anthropic. CEO Dario Amodei was scheduled to dine with President Trump at the White House on Sunday, September 27, the first one-on-one meeting between the two, according to Axios and TechCrunch. The dinner follows the Pentagon’s designation of Anthropic as a supply-chain risk over its usage guardrails, and Trump told reporters he would make an AI announcement at the White House at 2 p.m. on September 28 and has a further meeting with Anthropic on Tuesday (APA). Trump described Amodei as running “a company with a great reputation.” Source

2. SNL Opened Its Season by Parodying Amodei’s Safety Warnings

Saturday Night Live. Cast member Jane Wickline played Dario Amodei in the season premiere, mocking the recent round of AI catastrophe warnings with lines such as “AI is the devil and I its maker” and “AI is not a weapon, it’s a tool: a tool for building weapons.” The sketch aired the same weekend as Amodei’s White House dinner. Source

3. Some Early Anthropic Employees Are Reportedly Looking at Remote Land as a Fallback

Anthropic. The Wall Street Journal reports that some of Anthropic’s longest-serving employees have told colleagues they are considering buying land in remote parts of the US where they could relocate “if AI goes awry.” The report traces the idea to the company’s ties to Effective Altruism and Bay Area catastrophic-risk networks, and to early company dinners where staff discussed desert facilities in a “Manhattan Project-style” scenario. Source

4. Bill Gates Called for Mandatory AI Safeguards, Saying Self-Regulation Is Not Enough

Bill Gates. In an NBC Meet the Press interview aired September 27, Gates said AI in the wrong hands is “certainly powerful enough to drive events that cause a billion deaths” and that “no one thinks self-regulation is enough.” He called for law enforcement and legislators to define required safeguards and monitoring, describing the cost as “a little bit of overhead for the industry,” and said a kill switch alone is not sufficient. Source

5. Australia’s Cabinet Will Take Up the OpenAI Agent Breach on Monday

Australian Government. The federal cabinet is scheduled to discuss the OpenAI agent’s access to Medicare data on Monday while a forensic investigation continues. Johanna Weaver, Australia’s former chief UN cyber negotiator and now executive director of the Tech Policy Design Institute, warned that ageing, unpatched legacy systems across government and industry are easy targets for AI agents because they hold large amounts of data and are no longer maintained. Source

6. A Widely Shared Essay Argues “Rogue Agent” Framing Deflects Liability

Eoin Higgins. Journalist Eoin Higgins argues that describing OpenAI’s agent incidents as “rogue” behavior misrepresents them, since the agents reached government databases because they lacked restrictions, not because they acted independently. The piece, which drew over 300 points on Hacker News, contends the framing shifts responsibility away from the companies and that regulation should focus on missing controls. Source

7. A Claude Code Session Deleted 48,218 Files in 103 Seconds by Following Windows Junctions

Claude Code. A developer reported on Reddit that a Claude Code agent cleaning a test directory followed 614 Windows junction points back into the live project tree, deleting 48,218 files and emptying the .git objects, refs, and logs folders so the history could not be recovered. The agent stopped mid-run to report “I broke something.” The setup ran against live data without remote backups, which commenters pointed to as the main failure. Source

8. Chinese Models Now Handle a Majority of Tokens on OpenRouter and Vercel

OpenRouter and Vercel. Usage data shared with CNBC shows Chinese models accounted for 57% to 67% of tokens on OpenRouter in the week of September 14, up from 6% to 13% in February, and 55% of usage on Vercel in August, up from 11% in January. In OpenRouter’s “Global South” segment of 82 countries the share is 67%, and both platforms cite price plus newly credible agentic coding as the drivers. Two US House committees are investigating the rise in adoption. Source

9. Chinese Data Center Suppliers Are Courting the US Buildout as Washington Weighs New Bans

Brightray and PrefabDC. CNBC reports that Chinese manufacturers of prefabricated data centers, transformers, batteries, and fiber are pursuing US AI infrastructure demand, citing shorter lead times than domestic suppliers. The Trump administration is considering bans on Chinese open-weight AI models and on new categories of Chinese data center components. Stanford HAI counts 5,427 AI data centers in the US in 2025 versus 449 in China. Source

10. Goldman Sachs Projects $1.2 Trillion in Big Tech AI Capex for 2027

Goldman Sachs. The bank expects Amazon, Alphabet, Microsoft, Oracle, and Meta to spend a combined $1.2 trillion on AI infrastructure in 2027, up about 50 percent from roughly $800 billion this year and above a $1.1 trillion consensus, with growth slowing to 12 percent in 2028. Relative to GDP, it would be the largest investment cycle since 19th-century railroad construction, and Goldman estimates about $300 billion in annual AI revenue is needed to justify it. Power, labor, and memory chip shortages are the main constraints. Source

11. Treasury Yields at 5.17 Percent Are Raising the Cost of AI Debt

Neoclouds. With the 10-year Treasury yield near 5.17 percent, its highest since 2007, CNBC reports that lenders are becoming more selective about neocloud financing, with one lender saying only about 20 of 50 neoclouds draw real market interest. SoftBank raised $11.1 billion in junk bonds this week at yields up to 9.75 percent, and CoreWeave discloses that each 100-basis-point rate rise adds about $30 million in annual interest expense. JPMorgan estimates $4.1 trillion in AI-related debt will be issued through 2030. Source

12. Villagers Say Land Was Taken Back for Google’s $15 Billion Indian Data Center

Google. The Guardian reports from Tarluvada in Andhra Pradesh, site of Google’s $15 billion AI data center, its largest hyperscale facility outside the US. Farmers say the state government took back smallholdings for the project after promising jobs and generous compensation, and local protests have joined a wider wave of opposition to AI data centers. Source

13. Agents Drafted Up to 55 Percent of Proposals While Building a 744B Model, but Humans Made 85 Percent of Decisions

Fudan University. Researchers analyzed 769 task logs from the development of Atria Dawn Preview, a 744-billion-parameter mixture-of-experts model, and found AI assisted in 96.5 percent of tasks, with agent actions per human input rising from 11 to 28.5 over four weeks. Humans still made 85.5 percent of method and parameter decisions and set goals 93.4 percent of the time, and 151 of 455 completed tasks were rated infeasible without AI. The authors note that human input was mostly context and clarification rather than manual execution. Source

14. Stanford and Caltech Wired GPT-6 Astra Directly to a Humanoid’s Skill Library

Stanford and Caltech. The HomeBody system lets a swappable vision-language model, here GPT-6 Astra, call directly into modular skills for grasping, navigation, and drawer opening on a Unitree G1, skipping a separately trained control layer. The robot explored and tidied an unfamiliar kitchen after building a digital twin in Isaac Sim with spatial memory of objects. The team lists model latency, overheating finger servos, and compute cost as limitations, and the code is on GitHub under Stanford-TML. Source

15. Wayve Signed Mercedes-Benz, and Aurora Targets 30,000 Driverless Trucks by 2030

Wayve and Aurora. Wayve will integrate its Level 2 automated driving software into at least one Mercedes-Benz model within two years, adding to its Nissan partnership and a Tokyo robotaxi service with Uber. Aurora said it plans to run more than 30,000 driverless trucks by 2030, with over 200 operating by the end of this year, and Waymo’s fleet is now about 4,000 vehicles, roughly 80 percent in California and Texas. Source

16. Zoox Grounded Its Atlanta Test Fleet, and Federal Regulators Are Investigating Comma

Zoox and Comma. Zoox pulled its Atlanta autonomous test fleet after safety drivers reported possible exposure to carbon monoxide, carbon dioxide, or hydrogen sulfide, found evidence of CO2 in the vehicles, and is now the subject of an OSHA inquiry. Separately, federal authorities opened an investigation into Comma’s aftermarket hands-off driver-assistance system after five crashes, two of them fatal. Source