AI companies sign a safety accord: what the agreement actually changes

On September 29, the White House announced an agreement with leading AI companies on safer technology development. Reuters confirms the document was signed. It sets out industry self-regulation commitments rather than creating a new international law.

Who joined and what they promised

Participants include Google, Anthropic, Meta, OpenAI, NVIDIA and Elon Musk. ABC News reports that the commitments are voluntary. A signature therefore does not establish that every participating company has already implemented every measure.

The published agreement describes four layers:

• Internal safeguards for developing and using models.
• Dedicated teams that check those safeguards and address deficiencies.
• Independent external auditors or evaluators.
• An independent board committee overseeing findings and remediation.

How to judge whether it works

The useful next step for an observer is to look for concrete results: who conducted an audit, what they could examine, which findings became public and what changed afterward. The list of signatories cannot answer those questions.

For example, a successful model review tells readers little about a separate failure in an agent’s access to company email. Assessing protection requires knowing which component was examined. That is our practical interpretation of the agreement’s structure, not a report that such an audit has already taken place.

For individual users, access rules remain a separate decision. Before connecting an AI service to files, messages or payments, specify which actions it may carry out independently. A company’s signature does not replace those settings. We covered the technical side of this problem in our article on NVIDIA Open Agent Safety Platform.