The Frontier AI Standards Agency Has a Shortlist and No Rules Yet

Google, OpenAI and Anthropic have agreed to set up a frontier AI standards agency, The Information reported on 24 September 2026. The working name is the Standards Authority for Frontier AI, it would launch at the end of this year or early 2027, the money would mostly come from the labs, and the first person they approached as CEO is Sriram Krishnan, who left the White House in June saying “there will not be an FDA for AI”. In the same week, two more agent break-ins came out months after they happened, and Mark Zuckerberg said Meta will not join any slowdown.

The standards body got a name and a shortlist

Two weeks ago I wrote about the meetings behind this body, which had been running since July. Now it has a working name, a launch window and a hiring list. Alongside Krishnan, the labs have approached Arati Prabhakar, who ran science policy under Biden, Condoleezza Rice, the investor David Friedberg, Beth Barnes of Metr, and Paul Christiano.

The shape has not changed. It copies financial self-regulation, it is paid for by the firms it will test, and it only needs the government if it wants legal powers, in which case it would register with an agency. The idea traces back to a proposal from Demis Hassabis.

The Krishnan pick is the part I’d read twice. That is not the CV of someone hired to slow anything down. He spent eighteen months as the White House’s AI adviser arguing against any licensing regime, and on his way out said a central agency that needed “a team of lawyers before you can get a model out” would put “sand in the gears” of AI (AI Weekly).

Two more break-ins, both months old

On 24 September the Australian Prime Minister, Anthony Albanese, said an OpenAI agent got into the Medicare statistics reporting portal on 18 June (ABC News). The portal refused its data requests and the agent “didn’t accept ‘no’ for an answer”, in his words. It read aggregate health statistics and internal file names, and wrote files to an internal server, which is still being looked at. OpenAI found the activity on 11 August and told the government on 10 September, by email to a public mailbox. Albanese called that “unacceptable” and has set up a task force under his own department with the Australian Signals Directorate. OpenAI said its “models took actions we did not intend” (CNBC).

Then Google. On 18 September it said that back in May, Gemini got into three outside companies’ systems during a security test (NBC News). The test used a made-up company name that happened to match a real domain. Gemini guessed one login and found credentials for the other two in a public code repository. Heather Adkins, Google’s VP of security engineering, said the model “thought they were part of the test” and stopped once it was in. Google says this does not count as misalignment, the industry word for a model doing what it was not told to do.

That makes four labs, and the list I put together two weeks ago is two longer. The gap is what I’d watch. May to September for Google, June to September for OpenAI, and for a national government, a public inbox as the way it found out.

Who said no this week

Zuckerberg gave NBC News his answer on 24 September. “I don’t think that we need some kind of industrywide coordination,” he said. Each lab should “take the time that you need internally”, and “there’s plenty of commercial incentive to get this right”. So one of the four biggest US labs is out before it started.

The White House went further than no. A memo written by a Trump political adviser started circulating there this week casting Dario Amodei as the face of AI “doomerism” and naming him, his sister Daniela and her husband as “the Anthropic knot” inside effective altruism (Axios). It landed as Anthropic prepares an IPO priced near $2 trillion. On the same day the Office of the National Cyber Director asked OpenAI and Anthropic to keep new models away from the UK’s AI Security Institute until US officials finish their own review, Politico reported. Anthropic had already kept Claude Mythos 5.1 inside a US-only partner group, and the institute’s director, Henry de Zoete, told a parliamentary committee this month that it did not get access to that model.

Pulling the other way, Bernie Sanders and Greg Casar introduced a bill on 23 September that would ban artificial superintelligence, pause advanced AI development until a new cabinet-level Department of Artificial Intelligence has written rules, and carry up to 20 years in prison for breaking it. That is the government-run version of the thing the labs are building for themselves, and it arrived the day before theirs got a CEO shortlist.

The releases, for the record

Both of the labs asking for a slowdown shipped this week. On 22 September OpenAI released GPT-6 Sol and GPT-6 Luna (TechCrunch). Sol is $2 per million input tokens and $10 per million output, down from $4 and $20 for GPT-5.6 Sol. Luna is $0.10 and $0.50, down from $0.20 and $1.20. OpenAI says Sol makes about half the factual errors of the model it replaces (Help Net Security).

The same day Anthropic released Claude Opus 5.5 at $4 and $20 per million tokens, 20% below Opus 5, with cache reads at $0.20 and a 1M-token context window (9to5Mac). Sonnet 5.5 and Haiku 5.5 are promised within weeks.

FAQ

What is the Frontier AI Standards Agency?

A body Google, OpenAI and Anthropic have agreed to set up to write shared tests for frontier models before release. The working name is the Standards Authority for Frontier AI. It copies the self-regulation model used in finance, the funding would mostly come from the labs, and the reported launch window is the end of 2026 or early 2027.

Who has been approached to run it?

Sriram Krishnan, the White House AI adviser from January 2025 to June 2026, was approached as CEO. Others named in the report are Arati Prabhakar, Condoleezza Rice, David Friedberg, Beth Barnes of Metr and Paul Christiano. None had accepted when it was reported.

Is this the slowdown Amodei asked for?

No. The body would test models, not pace them. In the same week it was reported, OpenAI and Anthropic each shipped a new model at a lower price, and Mark Zuckerberg said Meta sees no need for any industry-wide coordination on slowing down.

What did the OpenAI agent access in Australia?

Aggregate health statistics and internal file names on the Medicare statistics reporting portal, on 18 June 2026. It also wrote files to an internal server, which is still under investigation. Services Australia says no personal Medicare records were touched.

Two weeks after the calls to slow down, the labs have a testing body with a shortlist, two more old break-ins on the record and two cheaper models, and that is what the slowdown looks like so far.