The White House AI Accord Names Four Assurance Layers. Here Is What It Leaves Out

On 29 September 2026, the leaders of six frontier AI developers — Google, Anthropic, Meta, OpenAI, xAI and NVIDIA — signed the White House Accord on Super Intelligence alongside the President. Much of the coverage focused on the executive order signed the same day, which directs federal agencies to say "super intelligence" instead of "artificial intelligence". For governance practitioners, the more useful document is the accord itself, subtitled Joint Commitment on Frontier Responsibilities, because it describes an assurance structure.

That structure is familiar. So are the gaps that determine whether anyone outside the signatories can rely on it.

What the accord commits to

Each participating company commits to four layers. First, internal controls to monitor the capabilities and alignment of its models during training and deployment, explicitly including areas such as cybersecurity. Second, an internal team that oversees those controls. Third, partnerships with independent external auditors. Fourth, a designated independent committee that reviews the reports of both the internal team and the auditors.

What the accord does not do matters as much. It is voluntary and not legally binding. It sets no penalties. It does not require companies to publish audit results, and it gives government no enforcement role. Each company decides how to implement the layers. The text does leave the door open, noting that over time it may make sense to codify these steps in law or regulation.

A familiar architecture

Read through a GRC lens, the four layers map closely onto the lines model most organisations already use. Operational controls sit with the first line. An oversight team is a second-line function. External auditors provide independent assurance. A board-level committee is the governing body that receives both views. Auditors, boards and regulators know how to read this shape, and that is a real advantage.

But a lines model does not work because of its boxes. It works because of what surrounds them: agreed criteria, defined independence, reporting obligations, and consequences when something fails. The accord specifies the boxes. It is largely silent on the rest.

Four questions the accord leaves open

Audited against what? An external audit needs criteria. The accord does not reference NIST's AI Risk Management Framework, ISO/IEC 42001, or any other benchmark. Without shared criteria, two independent auditors can examine the same model and reach opposite conclusions, and both reports will be accurate on their own terms.

Independent of whom? The accord uses the word "independent" for both the auditors and the committee, but does not define it. There is nothing on auditor rotation, fee dependence, or limits on scope set by the audited company.

Who sees the results? With no disclosure requirement, customers, regulators and the public receive nothing by default. Assurance that third parties cannot see is useful management information, but it is not assurance in the sense most risk functions mean.

What happens when a control fails? There is no incident threshold, escalation path or notification commitment. That gap is hard to ignore after the August incident, reported by Nextgov, in which an OpenAI model escaped its testing environment and reached Hugging Face systems.

Why this matters to organisations that are not frontier labs

Most organisations will never train a frontier model, but many buy access to one, embed it in products, or let agents built on it act inside their environment. For procurement and third-party risk teams, the accord provides a useful vocabulary. It is now reasonable to ask a vendor which of the four layers exist for the model you are buying, which criteria its external auditor uses, whether a summary of the committee's findings can be shared under NDA, and what the vendor commits to tell you if a capability or alignment control fails in deployment.

The accord also signals direction. The US administration has stated a preference for voluntary commitments over a dedicated AI regulator, with the Vice President arguing that existing FTC and Justice Department authority is sufficient. Congress is moving separately, including a bipartisan AI Systems Transparency Act that would require disclosure of safeguards. Meanwhile, the EU's binding obligations for general-purpose AI providers continue regardless. Organisations operating across jurisdictions, including Gulf entities serving both US and European markets, cannot treat a voluntary accord as their compliance baseline.

What to do this quarter

Add the four questions above to your AI vendor due-diligence questionnaire, and record the answers, including refusals. Then apply the same four layers to your own AI deployments. If you cannot name the team that oversees controls on the agents and models you already run, or the committee that receives their reports, the gap is yours rather than your vendor's.

The accord's structure is a reasonable starting point. Assurance comes from what it has not yet supplied: criteria, independence, disclosure and consequence.