AI News September 26, 2026: Google, OpenAI and Anthropic Anoint Their Own Regulator — SAFA Courts the Man Who Said 'No FDA for AI', Anthropic Exposes 200 Million Exchange Distillation Machine, OpenAI Rewrites GPT-6 Astra's Safety Record and Name-Drops 'Luna', the Labs Split on Going Public, and DevDay Looms Over Everything

If Thursday was the day AI governance turned political, September 26, 2026 is the day the labs stopped waiting for government entirely. The three companies that spent the summer pleading before the UN Security Council are now building their own regulator — staffed, ironically, by the people who once argued against one. Add an industrial-scale model-theft exposé, a quietly rewritten safety record for the world’s most powerful model, and a Tuesday keynote that could reset the developer platform, and you have one of the densest news weekends of the AI year. Here are the five stories that matter.

1. SAFA Takes Shape: The Labs’ Private Regulator Gets a Name and a Shortlist

Google, OpenAI, and Anthropic are moving to launch the Standards Authority for Frontier AI (SAFA) by late 2026 or early 2027, The Information reported this week. The self-regulatory body would set practical safety benchmarks, verify the labs’ public safety commitments, and may even run model assessments directly before commercial deployment — all independent of the US government.

The CEO shortlist is the story. Front-runner Sriram Krishnan served as Senior White House AI Policy Advisor from January 2025 until June 2026 — and left after publicly rejecting a federal licensing regime, declaring “there will not be an FDA for AI.” Former OSTP director Arati Prabhakar is also in the frame, with names like Condoleezza Rice and David Friedberg floating around the board conversation. The labs want the man who argued against a government regulator to run the private one instead.

The structure tracks Google DeepMind chair Demis Hassabis’s July proposal for a FINRA-style public-private body. OpenAI’s Chris Lehane framed it carefully: industry standards would “complement — not replace — mandatory federal safeguards.” Not everyone is convinced — Cohere CEO Aidan Gomez called the arrangement “a cartel by any other name” (CNBC, Sovereign Magazine).

2. Anthropic’s Threat Report: 200 Million Stolen Exchanges and a Named-Name Distillation Economy

Anthropic’s fourth threat-intelligence report — 154 pages covering December 2025 through August 2026 — reads less like a security bulletin and more like an indictment. The headline: illicit distillation at industrial scale, with the company naming Chinese AI labs directly.

The attributed volumes are staggering: Alibaba over 151 million exchanges (May–July 2026), Moonshot over 23 million, DeepSeek over 12.1 million in just 14 days of July, Zhipu over 3.4 million across 17 days, and Xiaomi more than 400,000 requests across 1,500+ accounts — in Xiaomi’s case, replaying saved customer coding conversations to manufacture training data. SenseTime allegedly bought harvested user-Claude conversations from intermediaries and used Claude itself to build the distillation pipelines.

One detail deserves every developer’s attention: when Anthropic’s cyber safeguards degraded Zhipu’s extraction attacks against Claude Fable, Zhipu’s employees simply switched to the older Opus 4.6 and another US lab’s model — capability theft routed around defenses like water finding a crack. None of the named labs have publicly responded (Anthropic, Daniel Miessler).

3. OpenAI Rewrites GPT-6 Astra’s Record — and Hints at ‘Luna’

Three weeks after launch, OpenAI’s September 22 system card update for GPT-6 Astra did something unusual: it corrected itself. HealthBench and HealthBench Professional scores were revised after OpenAI discovered “a misconfiguration in the previous evaluations,” and the alignment suite — including ExploitGym Honeypot and Coding Deception — was re-run with updated numbers.

The update also added an appendix covering GPT-6 Sol and GPT-6 Luna — the first official acknowledgment of a third “Luna” member of the GPT-6 family. With Astra posting 99.9% on ARC-AGI-3, 100% on ExploitBench, and a 3.4% misaligned-outcome rate (versus 18.8% for Sol), the family portrait is getting crowded — and the corrected record is a small but real win for evaluation transparency (OpenAI).

4. The IPO Divergence: Anthropic Races, OpenAI Waits

Reuters reports Anthropic’s annualized revenue run rate passed $65 billion by end of July — up from roughly $9 billion at the end of 2025 — with internal projections around $190–200 billion for 2028. A IPO marketing window could open as early as mid-October, though sources say the listing may slide past November’s US midterms.

Sam Altman drew the opposite conclusion: OpenAI will not go public in 2026, with Altman saying safety concerns make this “an ill-advised time” for a listing. One industry, two bets — Anthropic wants public capital while its growth curve is vertical; OpenAI wants freedom of action while the frontier is this unstable (Reuters).

5. Everything Points to Tuesday: DevDay 2026

OpenAI DevDay lands Tuesday, September 29 at Fort Mason in San Francisco — the first DevDay with global “Exchanges” satellite events in Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City. No announcements are confirmed, but with the GPT-6 generation three weeks old, SAFA negotiations behind the scenes, and the agentic platform wars raging between Amazon and Meta, expectations run to new models, agentic APIs, and developer tooling. The keynote is livestreamed free — set your alarm (OpenAI).

Sources: CNBC, ET Enterprise AI, Sovereign Magazine, AI Weekly, Anthropic, Daniel Miessler, OpenAI, Reuters.