ai-compliancetrade-pressNewsThe Broadside2 min read

Auditor flags gap in White House AI accord

The hard part isn't promising an outside audit, it's making outside mean more than a vendor label.


TL;DR

In a Federal News Network commentary, Patrick Sullivan, an A-LIGN executive who audits security and AI controls, says the voluntary White House Accord on Super Intelligence commits Google, OpenAI, Anthropic, Meta, xAI and Nvidia to outside reviews without defining auditors' rules. The immediate audience is the companies choosing reviewers and the firms competing to audit them. His timing warning for Congress is simple: delay lets convenience draft the definition.

Patrick Sullivan's argument is narrow, and that's why it matters. In a Federal News Network commentary, the A-LIGN executive says the Sept. 29 White House Accord on Super Intelligence gets the easy part right: Google, OpenAI, Anthropic, Meta, xAI and Nvidia voluntarily agreed to bring in an independent outside auditor to check whether AI safety controls work. His complaint is that the accord leaves the auditor undefined, even though the voluntary commitments could later be written into law.

That leaves a large hole inside a small word. Independence could mean a technically qualified firm with real system access. It could also mean a consultant already selling related services and reviewing management's slide deck. Sullivan says both could announce a passed outside audit under the accord's current language, even though the claims would mean different things.

His disclosed stake is part of the story. Sullivan says his firm audits security and AI controls and would likely compete for this work. He still argues rules should come first, including limits on other services, proof of expertise, enough access to test safeguards in practice and review of the auditor's work. The accounting analogy is the Public Company Accounting Oversight Board, which Congress created after early-2000s accounting scandals to inspect audit firms and enforce independence rules.

The concession is real. Sullivan agrees lawmakers shouldn't freeze model-testing techniques while frontier AI testing is still young. His distinction is useful: methods can evolve while baseline auditor qualifications are set. He points to existing international requirements for AI audit organizations. Connecticut, he notes, has created a state-supervised pilot to approve up to five outside organizations to verify AI systems against safety standards, and he says he testified before the California Senate on auditor standards.

The affected universe in the piece is narrower than normal compliance coverage. It starts with the companies that signed and the auditors seeking credibility; it reaches Congress only if the voluntary commitments become law. Sullivan's timing warning is that once companies start naming reviewers, the market may settle on its own definition of independence. If Congress later adopts the accord, it may inherit that definition. That's the compliance failure mode: label first, standard later.


Published ·Deep Fathom