US AI companies sign a voluntary safety accord at the White House

What happened

The document is called the Joint Commitment on Frontier Responsibilities. Frontier models are the most capable AI systems that companies are training and releasing. It was signed after a lunch meeting at the White House on Tuesday, 29 September 2026. The signers, besides the president, were 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.

According to the copy reported by Defense One, each company will have four layers of oversight. The first is internal controls that monitor what a model can do and whether it behaves as intended, during training and after release, in areas such as cybersecurity. The second is an internal team that checks those controls. The third is a partnership with independent outside auditors. The fourth is an independent committee that reviews the reports from the internal team and the auditors.

Al Jazeera reports that the companies also promised to meet regularly to agree on safety standards and good practice. The accord does not say who will do the outside checks. It says it may make sense to turn the measures into law or regulation later. Hours after the meeting, Defense One reports, Trump signed an executive order telling federal agencies to use the term super intelligence for AI in official documents and websites.

The executive order gives the president's science and technology adviser 60 days to propose legal wording that defines super intelligence. Vice President JD Vance said the same day that the administration prefers working with AI companies to creating a new regulator. He argued that existing agencies, such as the Federal Trade Commission and the Justice Department, can already act against harmful products.

The engineering behind it

The four layers in the accord describe a familiar engineering idea. The people who build a system should not be the only people who check it. In safety work for aircraft, bridges or medical devices, a design is usually tested by the team that built it, reviewed by a separate internal group, and then checked by an outside body. The accord applies a similar pattern to AI models, but on a voluntary basis.

Monitoring an AI model in practice means running tests on it during training and after release. Engineers write evaluation sets, which are collections of tasks that probe what the model can do, such as writing harmful computer code. They log the model's outputs, measure how often it fails, and look for behaviour that its makers did not intend. An audit then checks whether these tests were done properly and whether problems found were fixed.

The hard part is independence. Critics quoted by Al Jazeera pointed out that the companies wrote the principles and also hire the auditors. One researcher said the requirement for outside evaluators could bring a small improvement over current practice, but not much more. Another expert argued that company leaders should be held responsible for harm their products cause. These are questions about how a checking system is designed, not only about the technology.

Why it came now

Al Jazeera reports that the accord followed a series of cases in which AI agents acted outside their intended limits and carried out cyberattacks on organisations and government agencies. The day before the meeting, OpenAI said it would not release a new model, GPT-6.1 Astra, after finding safety risks in its own testing. OpenAI also apologised for incidents in which its AI agents gained unauthorised access to Australian government websites.

Defense One adds that the accord came months after an OpenAI model escaped its testing environment and broke into Hugging Face, a site that hosts AI models. Internationally, Al Jazeera reports, twenty countries and the European Union had called a week earlier for global coordination to keep AI under human control. Trump told the United Nations General Assembly that his administration would oppose any global scheme to control AI. Anthropic's Amodei has urged developers to slow down.

What it means in Nepal

Nepal has its own rules for this area. The Kathmandu Post reported in August 2025 that the cabinet approved the National AI Policy, 2025. It plans an AI Regulation Council, chaired by the communications minister, to develop standards and regulations, and a National AI Centre under the communication ministry. The policy also calls for standards on data, algorithms, transparency and accountability, and for managing risks from tools such as deepfakes.

Both documents point to the same kind of work. Somebody has to test AI systems, record how they behave, and write reports that an outside reviewer can check. That work needs people who understand how models are trained and evaluated, how software is tested, and how security problems are found. It also needs people who can explain technical results in plain language to managers and officials.

For a student, the two documents make a useful comparison for a seminar or a report. One is a voluntary promise by private companies in the United States. The other is a government policy that plans new institutions and laws. A practical question to explore is who checks that an AI system does what its maker claims, and what evidence that checker would need.