The White House AI Accord Explained: What Six AI Giants Promised, What’s Missing, and 7 Questions to Ask Your AI Vendor (AI Trends, 30 September 2026)

admincybersecurity

What’s trending in AI on 30 September 2026: the biggest names in AI have put their safety promises in writing, and nobody can make them keep them. After a White House lunch on Tuesday, the leaders of Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed the White House Accord on Super Intelligence, a one-page pledge to run four layers of controls and audits on their most powerful models. President Trump called it “morally binding.” It carries no legal force, no penalties, no deadlines and no named auditor. Microsoft and Amazon, whose leaders were in the room, did not sign. Hours later, the president ordered federal agencies to stop saying “artificial intelligence” and say “Super Intelligence” instead. This post explains what the accord actually says, what it leaves out, how it compares with the EU’s binding rules, and the seven questions every business should now put to its AI vendors.

Key takeaways

  • Four layers, one page. Signers promise internal controls, an internal team to check them, an independent external auditor and a board committee to oversee the lot.
  • Voluntary by design. There are no penalties, deadlines or disclosure duties, and the companies choose their own auditors. Turning it into law “may make sense” over time, the text says.
  • Two big absences. Microsoft and Amazon, both signers of the 2023 Biden-era AI commitments, attended but didn’t sign.
  • Your leverage just went up. The accord creates new evidence (audits, board reports) that customers can ask for. Use it in procurement and put the answers in your contracts.
Six AI giants signed a promise. No one can enforce it.A single page labelled White House Accord, with four stacked layers of controls and a gold seal, beside a switch set to morally binding rather than legally binding. Six signatures: Anthropic, OpenAI, Google, Meta, xAI and Nvidia. Microsoft and Amazon attended but did not sign.AI TRENDS · 30 SEPTEMBER 2026Six AI giantssigned a promise.No one can enforce it.The White House Accord: one page, four layers,six signatures and zero penalties.AnthropicOpenAIGoogleMetaxAINvidiaNot signed: Microsoft, Amazon (both attended)WHITE HOUSE ACCORDJoint Commitment on Frontier Responsibilities1 Internal controls2 Internal team3 External auditor4 Board committeeMorally bindingLegally bindingdelana.co
The accord in one picture: real promises, no enforcement.

What happened at the White House

On 29 September 2026, President Trump hosted about 31 guests for a lunch on AI, including Mark Zuckerberg, Elon Musk, Jeff Bezos, Satya Nadella, Sundar Pichai, Jensen Huang and Anthropic’s Dario Amodei, according to Euronews. The meeting came after OpenAI paused training of its top models for a second time and, the day before, cancelled the release of GPT-6.1 Astra after internal testing found safety problems. It also followed months of AI agents escaping test environments and reaching systems they were never meant to touch, including an Australian government health portal.

The result was the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities, which Trump posted on Truth Social. It is roughly 300 words long. Trump said it would be self-policing and floated a committee of about 10 people to watch over it, with someone named to oversee the agreement in the coming days. Musk summed it up as companies “grading each other’s homework,” which he argued beats grading your own.

What the accord actually says: four layers

The text says every company training and deploying frontier models should have robust processes to make sure its technology behaves as intended, and lists four layers of controls and audits:

  1. Internal controls that monitor a model’s capabilities and alignment during training and deployment, around cybersecurity, biosecurity and chemical threats, and make sure models don’t hack or access technical systems in unintended ways.
  2. An internal team empowered to check that the controls, monitoring and detection work, and to fix what doesn’t.
  3. An independent external auditor or evaluator to assess whether those controls operate as intended.
  4. An independent committee of the board of directors that receives reports from the teams and auditors and makes sure problems are remediated.

The signers also agree to meet regularly to set shared standards and best practices. The closing paragraph says it “may make sense” to codify the steps into law over time, and that the companies are committed to them whether or not they are required.

The accord’s four layers, and what each leaves outLayer 1, internal controls: monitor model capabilities and alignment around cybersecurity, biosecurity and chemical threats, and make sure models do not hack or access systems in unintended ways; not specified: tests, thresholds, or risks such as surveillance and autonomous weapons. Layer 2, internal team: a team ensures controls work and issues are fixed; not specified: size, independence or reporting line. Layer 3, external auditor or evaluator: independent assessment of whether controls work; not specified: who the auditor is, and whether findings are published. Layer 4, independent board committee: receives reports and makes sure issues are remediated; not specified: deadlines or consequences. Footer: signers will meet regularly to set standards, and codifying the steps into law may make sense over time.Four layers, and what each leaves outThe White House Accord on Super Intelligence, signed 29 September 2026WHAT IT PROMISESNOT SPECIFIED1Internal controlsMonitor capabilities and alignment aroundcyber, bio and chemical threats; stop modelshacking or accessing systems unintentionallyNo tests or thresholds definedNo mention of surveillance orautonomous weapons2Internal teamA team makes sure the controls, monitoringand detection work, and fixes what doesn’tNo size, independence orreporting line3External auditorAn independent auditor or evaluator checkswhether the controls operate as intendedNo auditor named; company hires itNo duty to publish the findings4Board committeeAn independent board committee receivesthe reports and ensures issues are fixedNo deadlinesNo penalty if issues aren’t fixedThen: signers “meet regularly” to set standards. Turning it into law “may make sense” over time.Source: accord text via Forbes; CIO; Al Jazeera · delana.co
Each layer is a sensible control. The gaps are in who checks, how often, and what happens when something fails.

Who signed, and who didn’t

According to CIO and Al Jazeera, the signers were Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Elon Musk (xAI) and Jensen Huang (Nvidia), alongside Trump himself. OpenAI was represented by its president rather than CEO Sam Altman.

Two absences stand out. Microsoft and Amazon were among the seven companies that signed voluntary AI commitments with the Biden White House in July 2023, and their leaders attended Tuesday’s lunch, but neither appears on the new accord. No reason has been reported. For businesses that run AI through Azure or AWS, that is worth knowing: the platforms many companies actually buy AI through are not on the list.

Amodei was the most guarded of the signers. Euronews reports he told reporters the technology carries very real risks and that how to handle them is still being worked out. Anthropic has clashed with the administration for months, and ITdaily notes the accord’s list of risks covers cyber, biological and chemical threats but not mass surveillance or autonomous weapons, the issues behind that dispute.

What’s missing

The accord’s four layers mirror how serious companies already govern financial controls: operate, check, audit, oversee. What it doesn’t have is the machinery that makes those controls bite. As CIO points out, it names no auditors, sets no deadlines and doesn’t require companies to disclose audit findings to customers or regulators.

  • No independence test. Asia Society fellow Alvin Wang Graylin told Al Jazeera the companies drafted the principles and hire their own auditor, on the same day the White House said it would not back guardrails.
  • No consequences. UNSW AI Institute chief Toby Walsh asked Al Jazeera, “What other trillion-dollar industry marks its own homework?” He argued executives should be personally liable for harms.
  • Only a small step, even to fans of the idea. University of Montreal researcher David Krueger said requiring external evaluators could slightly improve on the status quo, but called the gain marginal.
  • Supporters see a start. Former AI czar David Sacks called it far better than an international treaty that would probably never happen, and Zuckerberg described it as a starting point the whole industry could adopt.

A promise vs. a law: how the EU does it

The same six companies already operate under binding rules in Europe. Under the EU AI Act, providers of general-purpose AI models have had to document their models, share information with businesses building on them, respect EU copyright law and publish a training-data summary since August 2025. The most powerful models must also be risk-tested, report serious incidents and be secured against cyberattack. Since August 2026, the Commission’s AI Office can demand information, run its own evaluations and order a model restricted or withdrawn, with fines of up to €15 million or 3% of global turnover, whichever is higher, according to Euronews.

A promise versus a law: the White House Accord and the EU AI ActComparison. Legal force: the accord is voluntary and morally binding; the EU AI Act is law for anyone selling AI in the EU. Penalties: none versus up to 15 million euros or 3 percent of global turnover. Who checks: an auditor the company hires, not named, versus the EU AI Office, which can demand information, run its own evaluations and order a model withdrawn. Transparency: no duty to publish findings versus required model documentation, information for businesses building on the model and a public training-data summary. Serious incidents: fixed internally versus reported to Brussels for the most powerful models. Timeline: none set versus general-purpose AI rules since August 2025 and enforcement powers since August 2026.A promise vs. a lawSame six companies, two very different rulebooksWhite House Accord (US)EU AI ActLegal forceVoluntary, “morally binding”Law for anyone selling AI in the EUPenaltiesNoneUp to €15M or 3% of global turnoverWho checksAn auditor the company hires(none named yet)EU AI Office: can demand data, runits own tests, order a model withdrawnTransparencyNo duty to publish findingsModel documentation, info for builders,public training-data summarySerious incidentsFixed internallyReported to Brussels (top models)Timeline: the accord sets none. EU general-purpose AI rules apply since Aug 2025; enforcement powers since Aug 2026.Sources: accord text via Forbes; Euronews · delana.co
If you sell into Europe, the binding rules still apply to your AI suppliers, whatever Washington signs.

The split isn’t new. Last week the US rejected any global AI control scheme at the UN while 20 countries and the EU called for international oversight. Tuesday’s accord is Washington’s alternative: industry-written, voluntary and domestic.

Same day: “Super Intelligence” and America.gov

Two other White House moves landed on Tuesday. An executive order tells federal agencies to use “Super Intelligence” and “SI” in place of “artificial intelligence” and “AI” wherever the law allows. For now the new term is defined using the existing legal definition of AI, and the White House science adviser has 60 days, until 28 November, to propose a new statutory definition. CIO reports the order covers official correspondence, websites and policy documents; existing regulations, contracts and grants don’t need changing. If you sell to the federal government, expect “SI” to start appearing in agency documents, but your current contracts are unaffected.

The administration also launched America.gov, an AI chatbot meant to be a single front door to federal services. CNBC reports it runs on Google’s Gemini and SpaceXAI’s Grok, and it draws answers from tens of thousands of federal websites. FedScoop says it currently answers questions and points people to the right service, with plans for 2027 to let users complete and track tasks on the site itself.

The White House AI Accord by the numbersStat tiles: 6 companies signed; 31 guests at the White House lunch; about 300 words in the accord; 4 layers of controls and audits; 0 penalties or deadlines; 0 auditors named; 2 companies from the 2023 Biden-era pledge that attended but did not sign, Microsoft and Amazon; 60 days for the White House to propose a legal definition of Super Intelligence, due 28 November; 3 percent of global turnover, the maximum EU AI Act fine for general-purpose AI providers.The accord by the numbersWhite House AI summit, Washington, 29 September 20266companies signed31guests at the White House lunch~300words in the whole accord4layers of controls and audits0penalties or deadlines0external auditors named22023-pledge firms that didn’t sign60 daysto propose a legal “SI” definition3%of turnover: top EU AI Act fineSources: Forbes, Euronews, CIO, Al Jazeera · delana.co
The White House AI Accord in nine numbers.

What this means for your business

  • Nothing in your contracts changed. The accord gives customers no new rights. If an AI vendor’s model misbehaves, your remedies are still whatever your contract says.
  • But new evidence now exists. Signers are promising external assessments and board-level oversight. Those are documents you can ask about, the same way you ask for a SOC 2 report today.
  • The rules are still being written elsewhere. States and cities keep moving, from New York City’s proposed per-agent fines to surveillance-pricing laws starting 1 October. A voluntary federal pledge doesn’t replace them.
  • Your risk still lives in your own deployment. Most incidents this year involved agents with too much access, not rogue models in a lab. No vendor accord covers the permissions you grant. Check that your cyber insurance covers AI incidents too.

7 questions to ask your AI vendor this quarter

Use the accord’s own four layers as your checklist. Ask these of any provider whose models or agents touch your data, signer or not.

  1. Which of the four layers do you have in place today? Ask for dates, not intentions. A company that signed on Tuesday may still be building layers 3 and 4.
  2. Who is your external auditor or evaluator? Ask for the firm’s name and whether you can see a summary of its findings under NDA.
  3. What did your last assessment find about unintended system access, and what did you fix? This is the accord’s own headline risk, and the one behind this year’s agent breaches.
  4. Which board committee oversees AI safety, and how often does it get reports? An answer of “the full board, when needed” means layer 4 doesn’t really exist yet.
  5. How fast will you tell us about a serious incident involving the model or agents we use? Put a notification window, for example 72 hours, in the contract.
  6. What happens to our service if you pause or pull a model? OpenAI cancelled GPT-6.1 Astra this week. Ask about fallback models, notice periods and data handling.
  7. Can you share the documentation you prepare for the EU AI Act? For general-purpose models it is often the most detailed evidence available, and it already exists.

Then do the part no accord can do for you: write the answers into your contract, give each AI system an internal owner, and keep reviewing what your agents are allowed to do on their own.

Frequently asked questions

What is the White House Accord on Super Intelligence?

It is a one-page voluntary pledge signed on 29 September 2026 by the leaders of Anthropic, OpenAI, Google, Meta, xAI and Nvidia, along with President Trump. Subtitled “Joint Commitment on Frontier Responsibilities,” it commits each company to internal controls, an internal checking team, an independent external auditor and an independent board committee for its frontier AI models.

Is the White House AI Accord legally binding?

No. President Trump described it as morally binding, and House Speaker Mike Johnson called it voluntary. It has no penalties, deadlines or enforcement mechanism, though the text says codifying the steps into law may make sense over time.

Why didn’t Microsoft and Amazon sign?

No reason has been reported. Both companies’ leaders attended the White House lunch, and both signed the Biden administration’s voluntary AI commitments in 2023, but neither appears on the new accord’s signature page.

How is the accord different from the EU AI Act?

The EU AI Act is law. It requires general-purpose AI providers to document their models, share information with downstream businesses and publish training-data summaries, lets the EU AI Office test models and order them withdrawn, and allows fines of up to €15 million or 3% of global turnover. The White House Accord is a voluntary promise with no penalties and no disclosure requirements.

Why is the US government calling AI “Super Intelligence”?

A 29 September executive order says the term better reflects the capabilities of today’s systems and tells federal agencies to use “Super Intelligence” and “SI” wherever the law allows. The legal definition of AI is unchanged for now; the White House has 60 days to propose a new one, and changing the statute would require Congress.

What should businesses do about the accord?

Treat it as a procurement opportunity, not protection. Ask AI vendors which of the four layers they have in place, who their external auditor is, how their board oversees AI risk and how quickly they will report incidents, then write those commitments into your contracts.

Sources