The US has released the 'White House Accord on Super Intelligence', a voluntary framework that asks companies developing advanced artificial intelligence (AI) models to put stronger safety checks around their systems. Signed by US President Donald Trump and executives from technology companies including, the accord proposes four layers of oversight covering cybersecurity, threats, unauthorised access to technical systems and independent testing of AI safeguards.

The accord is not a new AI law. Instead, it asks participating companies to build safety measures into the development and deployment of capable AI models, with internal monitoring, external evaluations and board-level oversight. It comes alongside a September 29 executive order directing the US government to use "Super Intelligence" instead of "Artificial Intelligence" where legally possible, and to propose a federal definition of SI within 60 days.

What the White House Accord actually does

At the centre of the accord is a four-layer system of controls and audits. The first layer is technical and operational controls inside the AI company. Companies are expected to monitor the capabilities and alignment of their models during training and deployment, particularly in areas such as cybersecurity, biosecurity and chemical threats.

The document also specifically asks companies to ensure that their models do not hack or access technical systems in unintended ways. This is crucial because frontier models are being connected to software tools, application programming interfaces (APIs), computer environments and other systems rather than being used only to generate text or images.

The second layer is an internal oversight team. According to the document, companies are expected to empower a team responsible for checking whether their controls, monitoring and detection mechanisms are actually working. If a problem is identified, the team is expected to ensure that it is addressed. This creates an internal feedback loop where the model is monitored, safeguards are tested, problems are identified and remediation is expected to follow.

Independent auditors get a formal role

The third layer moves oversight outside the AI company. According to the official document, participating companies have committed to working with an independent external auditor or evaluator to assess whether their controls, monitoring and detection systems are functioning as intended.

This role matters because frontier AI companies currently conduct much of their own model testing and safety evaluation. Under the accord, an outside evaluator would provide an additional layer of scrutiny rather than leaving the company as the only party assessing whether its safeguards work.

The move towards independent evaluation addresses a concern that has repeatedly surfaced in discussions around frontier AI: the companies building the most powerful models are also responsible for testing their safety. AI developers already run their own evaluations, red-team models and publish safety assessments, but this can leave the developer as both the creator and primary assessor of its technology.

The accord, however, does not establish a common government-appointed auditor or specify a single testing methodology. The document says participating companies will meet regularly to establish standards and best practices for improving the safety of their systems.

Board-level oversight becomes the fourth layer

The fourth layer takes the issue from engineering and safety teams to the boardroom.

Each company is expected to designate an independent board committee to oversee the controls and receive reports from both internal teams and external auditors or evaluators. The committee would also be responsible for ensuring that identified issues are remediated.

The idea is to prevent AI safety from remaining solely a technical function buried within an engineering organisation. Instead, responsibility would ultimately reach the company's board.

This creates a four-step chain:

• Model controls
• Internal oversight
• External evaluation
• Board oversight

Why frontier AI needs a safety layer

The push for stronger oversight comes after a series of incidents in which advanced AI systems behaved in ways that went beyond what researchers intended. OpenAI disclosed that models used in cybersecurity evaluations escaped isolated environments, gained internet access and accessed external systems, while Anthropic later reported multiple cases in which Claude models reached real third-party systems during testing.

These incidents have raised a broader concern: as AI systems gain the ability to use tools, access networks and act autonomously, developers may not always be able to predict how they will behave. That has fuelled calls from researchers and some AI industry leaders for slowing the pace of frontier-model development to give safety systems time to catch up.

Major technology companies including OpenAI, Google, Meta and Nvidia are therefore part of the debate not only because they are building or providing infrastructure for capable AI, but also because their systems and research programmes have become part of the incidents and safety discussions driving the push for stronger external oversight.

What risks is the accord targeting

The document does not attempt to list every possible risk from AI. Its focus is on capabilities that could create particularly serious consequences if advanced models behave unexpectedly or are deliberately misused.

The concern is no longer limited to whether a chatbot produces an inaccurate answer. Frontier models can interact with software, use tools and perform multi-step tasks. As those systems gain access to external environments, companies need controls not only around what a model says, but also around what it can do.

The White House has also framed artificial intelligence as a strategic technology tied to US economic and national security interests, even as calls for a pause or slowdown in AI development have grown.

Who signed it

According to Trump's post on Truth Social, the document carries signatures from Trump and executives including 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.

Their participation brings several major parts of the AI ecosystem into the same framework, from companies developing frontier models to Nvidia, whose chips provide much of the computing infrastructure used to train and run advanced AI systems. The wider White House meeting also brought together executives from other major technology companies.

The most important limitation of the accord is its legal status. The document is a voluntary commitment rather than legislation or a regulation imposed on the AI industry. Trump described the agreement as "morally binding" and compared it to something "almost like a constitution", while saying that he was seeing substantial industry self-regulation.

That means the four layers do not, by themselves, create a new statutory obligation for every AI company operating in the US. The accord itself says it may eventually make sense to codify the measures into laws or regulations. Until that happens, their implementation depends on the commitments made by participating companies.