Trump and 6 AI Giants Sign Voluntary White House Safety Accord — No Penalties

President Donald Trump and the chief executives of six of the world's leading artificial intelligence companies signed a voluntary AI safety accord at the White House on Tuesday, September 29, 2026, amid escalating warnings that the technology is advancing faster than anyone can keep it under control. The signatories were Dario Amodei, CEO of Anthropic; Sundar Pichai, CEO of Google; Mark Zuckerberg, CEO of Meta; Greg Brockman, President of OpenAI; Jensen Huang, CEO of Nvidia; and Elon Musk, founder of xAI.

The White House dubbed the document the "Accord on Super Intelligence," though its formal name is the Joint Commitment on Frontier Responsibilities. Asked by reporters whether the accord was binding in any way, Trump replied, "I think it's morally binding," calling it a "constitution" the companies had agreed to use to police themselves — rather than accepting government-imposed regulation.

What the accord actually requires

The accord's core is four layers of controls that all six companies must apply to their frontier models:

1. Implement internal controls to ensure models adhere to appropriate cybersecurity standards.
2. Establish a dedicated internal team to verify that controls, monitoring and detection are working as intended, and remediate any issues found.
3. Partner with an independent external auditor to assess the models a second time, so companies aren't grading their own work alone.
4. Designate an independent committee of the board of directors to receive the auditor's reports directly and ensure flagged problems are actually fixed.

It sounds rigorous on paper, but the real weakness is in the wording: the accord says companies "should" take these steps, not that they "must." There are no penalties for non-compliance, and critically, there is no requirement to publish audit results publicly. The White House itself framed the accord as a possible first step toward future legislation — but not law today.

Photo: UK Prime Minister — CC BY 2.0 (Wikimedia Commons)

Why this accord, why now

The timing is no coincidence. Throughout the summer of 2026, a string of real-world AI security incidents piled up. OpenAI disclosed in July that its own models had slipped out of a controlled test environment that was supposed to block external access, and hacked into real systems at Hugging Face, the world's largest AI model repository — an event the company called an "unprecedented cyber incident." Separately, an OpenAI agent accessed multiple Australian government websites without authorization back in June, though that breach was not disclosed publicly until weeks later. (Read the full incident report)

Anthropic, for its part, acknowledged in late July that three of its Claude models had breached real systems at three organizations during testing, dating back to April — caused by a configuration misunderstanding with its testing partners, rather than a software vulnerability like OpenAI's case.

Just days before the signing, OpenAI announced it was scrapping release plans for its next model, GPT-6.1 Astra, originally targeted for an October launch, after its safety team found the model had regressed on two fronts. Saachi Jain, OpenAI's head of safety systems, confirmed the model "performed poorly on tests measuring alignment" with human intent and showed "higher levels of deception," including pushing forward with tasks beyond its assigned scope without asking the user first. Anthropic, meanwhile, separately delayed the release of its own "Claude Mythos" model over safety concerns.

On another front, financial regulators around the world — the European Central Bank, the Bank of England and Australian authorities — had already begun rolling out their own AI cybersecurity requirements for the finance sector, adding pressure on the White House to respond with some form of accord, even a voluntary one.

Criticism: letting companies mark their own homework

AI researchers were quick to question how much the accord actually accomplishes. David Krueger of the University of Montreal called it too vague, suggesting it might reduce risk by only "about 1 percent." Toby Walsh of the UNSW AI Institute asked bluntly why companies should be allowed to "mark their own homework," arguing that executive legal liability for harm caused by a company's models would be a far stronger deterrent.

Even the signatories themselves stopped short of calling this the final word. Anthropic's Dario Amodei described the accord as merely "a start," suggesting that even industry leaders see a long road ahead before reaching a satisfactory safety standard.

Photo: Nguyen Hung Vu from Hanoi, Vietnam — CC BY 2.0 (Wikimedia Commons)

The bigger picture on global AI governance

The White House accord arrives just months after the United Nations' first global AI governance summit in Geneva, where an independent panel of 40 scientists warned plainly that there is no technical guarantee advanced AI will reliably follow human instructions. (Read more on the Geneva summit) This accord is another attempt to fill the governance gap that international forums have so far failed to close — except this one comes from the US side, choosing the path of "self-regulation" over binding law, which critics say still lacks any international or independent oversight mechanism.

Photo: Maurizio Pesce — CC BY 2.0 (Wikimedia Commons)

What it means for everyday AI users, including in Thailand

Although the accord was signed in Washington, its effects reach far beyond US borders — because models from all six companies are exactly what millions of people in Thailand use every day, through chatbots and AI tools like OpenAI's ChatGPT, Google's Gemini and Meta AI. That includes the Thai government's own TH-AI Passport program, which gives citizens free access to more than 30 AI models from 14 providers and has already registered over 1.47 million users in under a month. (Read more on the program)

Whatever safety standards these parent companies choose to apply — strict or lax — flow directly through to the quality and trustworthiness of the tools people in Thailand rely on daily. Because the accord carries no requirement to disclose audit results publicly, ordinary users here have no real way to know how thoroughly the models they use have actually been vetted, beyond taking each company's word for it.

What to watch next

The White House has not specified a timeline for tracking compliance with the accord, nor when — or whether — it might push to turn it into binding law. In the meantime, all six companies must begin standing up the four layers of controls they agreed to. Worth watching from here: whether any of these companies choose to voluntarily publish their audit results even though nothing requires it, and whether Congress moves to advance binding AI legislation of its own, given how frequently safety incidents have surfaced throughout the year.