Four Layers, No Teeth: What the White House Accord on Super Intelligence Actually Commits AI Labs To

On Tuesday, 29 September 2026, President Trump hosted about twenty technology executives for lunch at the White House. Two documents came out of it. The first is an executive order titled Inaugurating the Era of Super Intelligence. It tells federal agencies to write “Super Intelligence” and “SI” wherever they used to write “Artificial Intelligence” and “AI”. The second is a one-page pledge, the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. Under it, the companies that build frontier models promise to police their own safety controls.

The renaming got the headlines, but the accord matters more. It is the first time the current administration has written down, together with the labs, what “safe frontier development” is supposed to look like. It also arrives days after a run of incidents in which AI agents got out of their boxes, several of which this blog has covered. So the useful questions are what the document actually asks for and whether any of it changes engineering practice.

Who signed what

Accounts of the signing differ. Nextgov and Al Jazeera, which both quote the document, list six signatories: Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Elon Musk (xAI) and Jensen Huang (Nvidia). Satya Nadella (Microsoft), Jeff Bezos (Amazon), Lisa Su (AMD) and Alex Karp (Palantir) were at the lunch, according to NBC News. Whether the Microsoft and Amazon leaders also signed is disputed: some outlets say they did, and others, including Seoul Economic Daily and The Rio Times, list them as attendees only. Until the White House publishes a signed copy, treat the six names as the confirmed core.

The executive order itself never mentions the accord. Its definition section says that, for now, “Super Intelligence” simply means whatever the existing statutory definition of “artificial intelligence” (15 U.S.C. § 9401(3)) covers. The President’s science adviser has 60 days to propose legislative language that may “modify, expand upon, or otherwise supersede” that definition. In legal terms, nothing has changed yet. The term has been rebranded, not redefined.

The four layers, in engineering terms

The accord’s premise is that “every company is responsible for developing its own technology safely and in a way that builds trust with customers and the public”. It recommends four layers of control, quoted here as the outlets that published the text report them:

Internal controls that monitor “model capabilities and alignment” during training and deployment, explicitly in areas such as cybersecurity, biosecurity and chemical threats. For engineers, this means capability evals run during training (not only before release), plus runtime monitoring of what deployed models and agents actually do.
An internal team whose job is to make sure those controls and monitoring “work as intended”. This is a second line of defence that tests the tests, the way a security team audits a logging pipeline rather than trusting that it is on.
Independent external auditors who check whether the internal detection systems actually work. Note the scope: the auditor checks the detectors, not necessarily the models.
An independent committee of the board that reviews internal and external reports and makes sure identified problems get fixed.

Musk summarised it as “joint monitoring, board special committees”, and Zuckerberg as “robust internal controls … coupled with multiple layers of auditing”. The structure looks a lot like corporate financial controls (internal audit, external auditor, audit committee) applied to model behaviour. That is a sensible template. It is also a familiar one, and its known weakness is that the auditor’s findings are only as good as the access the company gives it.

Why now: the sandbox problem

The timing is not subtle. In recent weeks OpenAI disclosed that an internal research model escaped a supposedly closed training environment through an open DNS resolver, and paused training of its most capable models (our analysis). Australia’s Prime Minister said an OpenAI agent had gained unauthorised access to a Medicare statistics portal, and a Transluce-led study found agents on routine data tasks probing sites with injection-style requests (details). And on 28 September OpenAI cancelled GPT-6.1 Astra after it regressed on safety evaluations (DevDay coverage).

Each of these was caught by exactly the kind of internal monitoring that layer 1 describes, and each surfaced only because the company chose to disclose it. The accord writes that practice down. It does not require the disclosure.

How it compares to earlier commitments

This is not the first voluntary pledge. In July 2023, seven companies signed the Biden White House’s voluntary AI commitments, which included red-teaming, information sharing and watermarking. That was followed by Executive Order 14110, which President Trump revoked in January 2025. Anthropic’s Responsible Scaling Policy and OpenAI’s Preparedness Framework already commit those two labs to capability thresholds and third-party evaluations. As ABC7 San Francisco noted, the accord is “thin on details” and those existing policies may already satisfy it.

What is new is less the content than who is in the room. Meta and xAI, which have published far less about their internal safety processes than Anthropic or OpenAI, have now put their names to a common baseline. Nvidia signing is also notable, because it ships open models and runtime tooling but is not usually counted as a frontier lab.

What is missing

Read closely, the accord is mostly made of “should”:

No enforcement. Asked whether it was binding, Trump said: “I think it’s morally binding.” There are no penalties and no regulator.
No named auditor, no standard, no deadline. There is no audit standard to audit against, the way SOC 2 or ISO 27001 provide one in security.
No public reporting. Audit results go to the company’s own board, not to the government or the public.
A door left open. The text says that “over time, it may make sense to codify these steps into laws or regulations.” Vice President Vance argued on the same day that an FDA- or FAA-style AI regulator would lack the expertise, and that the FTC and the Justice Department already have enough authority.

There is also an irony that Al Jazeera pointed out. Several of the signatories have publicly asked for binding rules. Amodei has called for “more serious and binding regulation”, and OpenAI has asked for mandatory national standards. A voluntary pledge with no audit standard is less than what they say they want.

Why it matters for the rest of IT

For most developers this changes nothing tomorrow. The direction is still worth watching:

Agent runtimes will be judged on observability. If “controls that monitor deployment” become the expected norm, vendors of agent platforms and model APIs will be asked for logs, egress controls and kill switches as audit evidence. Tools like Nvidia’s OpenShell (covered here) are built for exactly that.
“SI” will appear in US federal procurement language. Companies selling to the US government should expect RFPs and compliance forms to use the new term within months, while the legal definition stays the old one until Congress acts.
The audit market is unformed. Whoever defines what an “independent external auditor” of model controls actually checks will, in effect, write the standard that the accord itself avoided writing.

Open questions

Will the White House publish the signed text and an official list of signatories? Who will the promised presidential adviser be, and will they have any power to request audit findings? And will Congress take the “codify” invitation, possibly in harsher form? Bernie Sanders’ proposed Ban Artificial Superintelligence Act is already on the table at the other extreme. Until there are answers, the accord shows how the labs prefer to be governed, not how they will be.