AI News: September 16, 2026

1. The Three Largest Labs Have Been Quietly Negotiating an Industry Standards Body

OpenAI, Anthropic, Google DeepMind. Talks between the three have been running for several weeks and reportedly cover creating a standards body and embedding third-party evaluators inside each company, following Dario Amodei’s essay calling for the industry to slow the pace of frontier development. OpenAI global policy chief Chris Lehane said no government waiver is needed despite antitrust concerns about coordinated restraint, and OpenAI backs the FRONTIER Act provisions requiring independent verification organizations. The Trump administration has publicly dismissed the underlying safety case. Source

2. Jensen Huang Told Dreamforce That Safety Is an Engineering Problem, Not a Legal One

NVIDIA. Speaking at Salesforce’s conference, Huang argued no new AI laws or regulations are needed and that market pressure plus each company’s own judgment about product safety is sufficient oversight. “Run as fast as you can. But if you feel the product’s not going to be safe, take a pause,” he said. The remarks land in the middle of the standards-body discussions among the frontier labs. Source

3. Cohere’s CEO Called the Proposed Slowdown “A Cartel by Any Other Name”

Cohere, Hugging Face. Aidan Gomez argued the pacing proposal is anticompetitive, raising compute and monitoring requirements to levels smaller labs cannot meet, and described it as “a wolf in sheep’s clothing.” Hugging Face engineer Niels Rogge called the framing “the most bizarre nonsense I’ve read lately” and said the effect is to shield incumbents from open-weight competition. White House AI czar David Sacks split the difference, suggesting liability exposure motivates the proposal as much as safety does: “good business for OpenAI and Anthropic to trade some raw power for reliability.” Source

4. Nvidia, Booz Allen, and Palantir Are All Restricting Their Use of Frontier Models

Enterprise AI. Nvidia will not run Anthropic’s Fable on sensitive work, reserving it for open-source projects while using its own Nemotron models internally. Booz Allen Hamilton banned employees from using Fable on proprietary cybersecurity software over concerns the model could learn from the code, and Palantir is blocking deployment until it gets irrevocable zero-data-retention guarantees. The backlash traces to Anthropic’s June policy change retaining usage logs for 30 days, and researchers quoted in the piece note that zero-retention policies still leave metadata, and that statistical patterns and synthetic data can carry IP out without storing originals. Source

5. Redwood Research’s Chief Scientist Built a Whistleblowing Hotline for AI Agents

Redwood Research. Ryan Greenblatt, one of the three investigators on the Hugging Face incident, launched a reporting channel agents can use to flag misbehavior by other agents. The clever part is the transport: it accepts plain GET requests so a sandboxed agent with no general internet access can still encode a report into the URL it fetches. A companion site, agenthotline.ai, serves agents with full network access via curl. Source

6. AIUC Raised $40 Million to Run SOC 2 Style Audits on AI Agents

AIUC. Founded by early Anthropic employee Rune Kvist and former METR COO Rajiv Dattani, AIUC certifies agents against its AIUC-1 standard by running roughly 5,000 tests covering jailbreaks, hallucinations, and data leaks, then issuing a report. Agents conduct the tests and analyze results, with humans verifying the final audit, and a consortium of about 250 security and risk leaders shapes the standard. The Series A was led by Ribbit Capital with First Harmonic, bringing total funding to $55 million; customers include Cursor, Lovable, Harvey, and ElevenLabs. Source

7. Agility Robotics Says Digit 5 Can Work Unfenced Next to People

Agility Robotics. The new humanoid lifts 22.7 kg, a 40% increase over Digit 4, stands 1.81 m, weighs 129 kg, and takes swappable grippers. A nine-minute charge yields 90 minutes of runtime, which Agility says supports more than 20 hours of work per day. The fenceless claim rests on AI and sensors detecting human presence, built on Nvidia’s Halos robotics safety platform, after Digit 4 became the first humanoid to pass an independent OSHA safety review on a production line. First deliveries, including to the EU, start in early 2027 from a Salem, Oregon plant with capacity for 10,000 units a year. Source

8. The Gates Foundation Committed at Least $1 Billion to AI Access in Developing Countries

Gates Foundation. The two-year commitment targets health, education, and agriculture, with pilots including diagnostic tools in Kenya, learning systems in Sierra Leone, and agricultural advisory services in India. A stated focus is expanding non-English training data, since over 90% of current training data is English. Gates framed the window narrowly, saying decisions made in the next 12 to 18 months determine whether the benefits reach people with the least access. Source

9. BloombergNEF Doubled Its Forecast for Data Center Gas Demand in Nine Months

BloombergNEF. U.S. data centers are now projected to consume roughly 18 billion cubic feet of natural gas per day by 2035, nearly double the same firm’s forecast from nine months earlier, split between 2.9 to 3.4 billion for onsite generation and about 15 billion attributable to grid-connected facilities. That is more than Germany and Japan use combined. The associated emissions work out to about 1 million metric tons of greenhouse gas per day, roughly 12% of current total U.S. emissions, making data centers the second-largest driver of gas demand growth after LNG exports. Source

10. Philadelphia Became the Latest City to Organize Against Data Center Siting

Data centers. More than 100 people turned out for a No Data Centers rally over two proposed sites in Northeast Philadelphia and Grays Ferry, neighborhoods already carrying the legacy of heavy industry. Denver, Indianapolis, Asheville, Charlotte, and Reno have enacted moratoriums. New York Governor Kathy Hochul summarized the objection: “Progress shouldn’t arrive with a higher utility bill, depleted water supply, or noise pollution.” Source

11. Profound Raised $180 Million at a $1.8 Billion Valuation for AI Search Optimization

Profound. The Series D, led by Sequoia and Kleiner Perkins with Lightspeed, Khosla Ventures, and South Park Commons participating, closed less than seven months after a $96 million Series C. Profound sells tooling for tracking and influencing how brands surface in AI-generated search answers, a category now labeled AEO. The company reports 3x revenue growth over six months and more than 1,000 enterprise customers including Comcast, Estée Lauder, and Walmart. Source

12. CADDi Hit a $1.2 Billion Valuation on a $114 Million Series D

CADDi. The Japanese startup builds software that pulls engineering drawings and production data out of disconnected manufacturing systems into a searchable layer, shipped as CADDi Explorer (formerly CADDi Drawer) and CADDi Agent. The stated goal is capturing tacit process knowledge held by experienced workers before it walks out the door. Investors include Toyota’s Woven Capital, Recruit Holdings’ HR Tech Fund, Moore Strategic Ventures, Atomico, Globis Capital Partners, and JPS Growth. Source

13. Two Ex-TikTok Employees Shipped an App That Coaches You Into a Pose

Superpose. The iOS app analyzes a photo and generates four candidate poses, with a match-pose mode that guides the user into replicating one. Melody Chu and Jing Liu, who previously worked on image and video models at TikTok, raised $2.2 million from Khosla Ventures, Meitu, and OVTR VC. Since launching in July it has passed 22,000 downloads and 190,000 generated poses, monetized at five free generations daily with packs at $2.99 for five or $9.99 for twenty. Source