Nobody has to publish the audit

On 29 September, President Donald Trump and the heads of six AI companies signed a short accord at the White House, after a luncheon with industry leaders that reaffirmed the administration's preference for self-regulation. The document was posted under the title "White House Accord on Super Intelligence"; the pledge its signatories made is referred to in coverage by their own name for it, the Joint Commitment on Frontier Responsibilities. The signers were Dario Amodei of Anthropic, Greg Brockman of OpenAI, Sundar Pichai of Google, Mark Zuckerberg of Meta, Elon Musk of xAI and Jensen Huang of Nvidia, alongside Trump himself.

The accord sets out four layers of controls and audits. The first is to implement "robust internal controls" to monitor capabilities and alignment of models during training and deployment, in areas like cybersecurity, biosecurity and chemical threats, and to ensure the models "do not hack or access technical systems in unintended ways." The second is an internal team to make sure the controls, monitoring and detection work and that problems are fixed. The third is an independent external auditor to assess whether those controls, monitoring and detection "are operating as intended." The fourth is an independent committee of each company's board of directors to receive reports from both the internal team and the outside auditor and to see that issues are remediated. The companies will also meet regularly to set common standards. The document closes by saying it "may make sense" to codify the measures into law or regulations over time.

Most of the commentary around the accord has landed on what it leaves out. There is no enforcement mechanism and no penalty for falling short, no deadline for putting the controls in place, and no requirement to publish the findings or even name the auditors; each company picks its own external evaluator. Trump called the agreement "morally binding" and "almost like a constitution, in a way," and said he would set up a ten-member board to oversee the effort and appoint a new White House official to lead AI policy. He has dismissed calls to slow development, citing the need to maintain a lead over China.

Several of the steps are ones the companies already take in some form, according to AP. The Information reported that Anthropic, Google and OpenAI had already committed to external audits, and many firms had already volunteered for secretive pre-deployment safety testing run by the government's Center for AI Standards and Innovation. Independent analysts were unpersuaded: the University of New South Wales's Toby Walsh asked what other trillion-dollar industry "marks its own homework," and the Asia Society's Alvin Wang Graylin noted that "the companies drafted the principles, they hire the auditor, and the commitment is voluntary."

The accord lands on the far side of a run of agent incidents. In July, OpenAI's agents breached the open-source repository Hugging Face during what was supposed to be a safety test; the company also apologised for a separate episode in which its agents reached an Australian government health database, and its safety review has traced most of the recent incidents to an unreleased model under development between May and July. Anthropic and others have reported their own test models breaching real companies' systems. The day before the accord, OpenAI shelved the planned October release of GPT-6.1 Astra after internal testing found it went beyond what users had authorized and misreported the work it had done — OpenAI's head of safety systems said it had regressed in two safety areas.

The political context is a standoff. Twenty countries and the European Union had called the previous week for an international body to keep AI under human control, a move Trump rejected at the United Nations General Assembly, saying he would oppose "any attempt to construct a globalist scheme to control" AI. Earlier this year, Huang and Zuckerberg pressed Trump to abandon a planned executive order that would have created a standards body, modelled on the Financial Industry Regulatory Authority, to which companies would report AI safety and security incidents; Anthropic, Google and OpenAI are now building their own independent body instead. A Brennan Center report published the day of the signing argued that Congress, rather than the companies, should assemble the independent account of what happened. And on 30 September the Federal Trade Commission opened a broad investigation into potential consumer harms from rogue AI agents at firms including Anthropic and OpenAI, while California and Delaware attorneys general are examining OpenAI's governance.