Confirmed: On 29 September 2026, President Donald Trump and the leaders of six companies — OpenAI, Google, Meta, Anthropic, Nvidia and xAI — signed a voluntary one-page document, the White House Accord on Super Intelligence, committing each company to four layers of AI safety controls and audits: internal controls, an internal checking team, an independent external auditor, and an independent board committee. Reported, and not established as fact: whether any company has completed an audit, or whether the arrangement will ever become law. Anthropic’s 15 October evaluator announcement is set out below; no audit findings had been published. The pact sets no deadline, names no auditor, and carries no penalty. This article explains each layer in the pact’s own terms and does not present the pact as law or as having penalties. The wider story of the signing is covered in the White House AI safety pact explainer.
What was signed on 29 September 2026
The document is formally titled the “White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities.” President Trump published the two-page text, including a page of signatures, on his social media platform late on 29 September, according to the Associated Press account of the signing. The signatories named on the accord are Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of xAI and Jensen Huang of Nvidia, alongside President Trump. House Speaker Mike Johnson was also present at the White House meeting, reporting of the event said.
The accord is voluntary. The Associated Press reported that President Trump called it “morally binding.” The text uses recommendatory language — each company “should” implement the four layers — and ends by saying that “over time, it may make sense to codify these steps into laws or regulations.” It does not say the steps are law now. CoinDesk’s coverage put the point plainly: the agreement “has no enforcement mechanism, disclosure requirement or implementation deadline, leaving companies to choose their auditors and address any shortcomings themselves.” The same day, President Trump signed a separate executive order directing federal agencies to use the term “super intelligence” instead of “artificial intelligence,” a change explained in the super-intelligence term explainer. That renaming is a communications instruction; it does not change what the accord requires of companies.
AI safety audits explained: the four layers — internal controls, external auditors, board oversight
The accord’s central passage says every company training and deploying frontier models should implement “four layers of controls and audits,” in addition to any other precautions. The four layers, in the accord’s order, are set out below, followed by a section on each one. The wording in quotation marks is the accord’s own.
| Layer | What the accord asks for | Who does it |
|---|---|---|
| 1. Internal controls | Monitor model capabilities and alignment during training and deployment, around cybersecurity, biosecurity and chemical threats; ensure models do not hack or access systems in unintended ways | The company itself |
| 2. Internal team | Ensure the controls, monitoring and detection are operating as intended, and that issues are remediated | A dedicated internal team |
| 3. External auditor | Independent assessments of whether the controls, monitoring and detection are operating as intended | An independent external auditor or evaluator |
| 4. Board committee | Oversee and receive reports from the internal team and the external auditor; ensure identified issues are remediated | An independent committee of the board of directors |
Layer 1: internal controls — what gets monitored
The first layer is the monitoring itself. The accord asks each company to implement “robust internal controls to monitor the capabilities and alignment of its models during training and deployment around areas like cybersecurity, biosecurity, and chemical threats.” Monitoring therefore covers training and deployment, capabilities — what a model can do — and alignment — whether it behaves as intended. CoinDesk reported that the pact “calls for monitoring advanced models for cyberattack, hacking and biological or chemical risks after experimental AI agents accessed unauthorized systems and AI was linked to cryptocurrency security breaches.”
Layer 2: the internal team that checks the controls
The second layer is an internal check on the first: the accord asks each company to “empower an internal team to ensure all of the controls, monitoring, and detection are operating as intended, and that any issues are remediated.” A monitoring system can fail in practice — missing a capability, misreading a result, or flagging a problem nobody fixes. The internal team’s job is to catch that gap and see problems fixed. The team reports, in the accord’s structure, to the board committee in layer 4, alongside the external auditor.
Layer 3: the independent external auditor — and the auditor-choice gap
The third layer is the one supporters describe as the most significant, and critics scrutinise most closely. The accord asks each company to “partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended.” An external auditor, in principle, answers the obvious objection to layers 1 and 2: that a company is marking its own work. An outside party with no stake in the result is supposed to test whether the controls genuinely work.
The gap is that the accord does not say who counts as independent, what qualifications an auditor needs, what criteria the auditor assesses against, or what happens if the verdict is unfavourable. Each company chooses its own auditor. CoinDesk reported that the pact “does not require companies to publish or name the auditors,” and Tech Times reported that it contains no requirement to use auditors the companies did not select themselves, and no requirement to report safety incidents to any government body. Audit findings go, under the accord’s structure, to the company’s own board committee — not to a regulator and not to the public. A reader therefore cannot tell from the accord whether an audit has happened, who conducted it, what it found, or whether anything changed, unless the company chooses to disclose those facts voluntarily.
On 15 October 2026 Anthropic announced it had begun giving employee-like access to a team from METR, that METR would publish its risk assessments, and that it was also working with SecureBio and Accenture as evaluators. No audit findings had been published, and no comparable appointment by another signatory was confirmed in primary sources checked.
Critics quoted in coverage made exactly that objection. USA TODAY quoted Mehta: “You basically are allowing companies to grade their own homework, and they might, of their own accord, be hard graders. You are really relying on the companies to implement this in a robust way.” Genevieve Smith, a research fellow at Stanford University and founding director of the Responsible AI Initiative at the University of California, Berkeley, called the agreement “insufficient” in the same report, saying it avoided the incentive structures encouraging ever more capable models while safety was deprioritised. Senator Elizabeth Warren was reported as responding: “Self-regulation? That’s a recipe for disaster.” Those are critics’ characterisations, reported in coverage checked for this draft, presented here as such — not as this article’s verdict on whether any audit under the pact will be rigorous.
Layer 4: board oversight — who sees the reports, and who must act
The accord requires the committee to be “independent,” but does not define independence, say how many members it must have, require any member to have AI safety expertise, or require the committee to disclose anything publicly. Its practical force therefore depends on its composition, access to raw findings, and willingness to delay or stop a deployment — none of which the accord specifies. The committee has no power beyond the company: it cannot fine the company, suspend a model, or refer a matter to a regulator. Its leverage is internal, through the board’s authority over management.
President Trump also said at the signing that about ten people would be named to a committee to “watch over the whole enterprise,” and that he would name someone to oversee the agreement in the coming days after consulting with industry, the Associated Press reported. That proposed committee is separate from the company board committees in layer 4. No members had been named in reporting checked for this draft, and the accord text does not create such a committee. Its status should be re-checked before publication, and no claim about its membership or powers is made here.
What the pact does not do: no deadline, no penalties, no publication
The most important limits are absences. First, there is no deadline: nothing says by when a company must have its controls, internal team, auditor or board committee in place, or by when a first audit must be completed. Second, there is no penalty for doing none of these things, or doing them poorly. Third, there is no publication requirement: nothing requires a company to name its auditor, publish audit findings, or report an incident to the government or the public. The accord’s only forward-looking commitment on enforceability is that it “may make sense” to codify the steps into law over time. Until that happens — if it happens — compliance rests on the companies’ own commitments and reputational pressure. The criticism of that structure is covered in the no-enforcement critics explainer. Supporters’ counter-argument, as reported, is that a written commitment from the companies themselves is a starting point that did not exist before in this form. This article takes no position between those views.
Frequently asked questions
What is an AI safety audit, in the pact’s sense?
An independent assessment by an outside auditor or evaluator of whether a company’s internal safety controls, monitoring and detection are operating as intended. The accord asks each company to partner with such an auditor, but does not define who qualifies, what standards apply, or whether findings must be published. It is a company-commissioned review, not a government inspection.
Who chooses the external auditor?
Each company chooses its own auditor. The accord says companies should “partner with an independent external auditor or evaluator,” but names no auditor, sets no qualification criteria, and creates no panel or regulator to appoint one. Reporting noted that companies are not required to name or publish their auditors. That choice is the central gap critics identify.
Is the White House AI safety pact legally binding?
No. The accord is voluntary, uses “should” rather than “must,” and carries no penalties. President Trump called it “morally binding.” The text says it may make sense to codify the steps into law or regulation over time, but no law has been created by the accord itself.
What is the deadline for companies to comply?
There is none in the accord. The pact sets no date by which controls, an internal team, an external auditor or a board committee must be in place, and no date for a first audit. No company timetable has been confirmed in reporting checked.
Updated: 1 November 2026. The accord text and signing details are as published on 29 September 2026; Anthropic’s 15 October evaluator announcement is noted in the body; no audit publication or enforcement action had been confirmed in reporting checked for this draft.
SOURCES
- American Presidency Project, White House Accord on Super Intelligence (full text) — https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence
- CoinDesk, AI safety pact report, 30 September 2026 — https://www.coindesk.com/tech/2026/09/30/openai-google-and-meta-pledge-outside-ai-audits-under-voluntary-white-house-deal
- InfoSecurity Magazine, Super Intelligence safety accord report, 30 September 2026 — https://www.infosecurity-magazine.com/news/trump-ai-giants-super-intelligence/
- Associated Press (via DTNPF), ‘Self-Police’ accord report, 30 September 2026 — https://www.dtnpf.com/agriculture/web/ag/news/world-policy/article/2026/09/30/trump-says-top-tech-firms-signed-ai
- USA TODAY, accord vs regulations explainer, 2 October 2026 — https://www.usatoday.com/story/tech/2026/10/02/trump-ai-ceo-self-police-accord-risks/92025139007/
- Tech Times, accord penalties and auditors report, 2 October 2026 — https://www.techtimes.com/articles/328464/20261002/white-house-ai-safety-accord-has-no-penalties-no-breach-reporting-self-chosen-auditors.htm
- Anthropic, “An update on independent evaluators”, 15 October 2026 — https://www.anthropic.com/news/an-update-on-independent-evaluators
