An Ars Technica article dated September 30 says that, in an agreement on Tuesday, about two dozen technology firms voluntarily committed to controls recommended by the White House. The commitments include independent safety audits meant to test whether internal controls, monitoring, and detection are actually working. The external reviews focus on risks related to cybersecurity, biosecurity, chemical threats, and unintended actions by models. The article names those review topics and does not describe audit procedures. The firms also agreed to meet regularly, discuss practices, and set common safety standards and benchmarks.

[1]
Graphite: a dense row of short ticks on a baseline, with an empty red circle at the right.
Short ticks crowd a baseline. An empty red circle sits at the right end of the same line. That is a signed page without a legal seal. An illustration, not the accord., AI-generated illustration, not a news photograph

Signers include Dario Amodei of Anthropic, Sam Altman of OpenAI, Elon Musk of SpaceXAI, Jensen Huang of Nvidia, Mark Zuckerberg of Meta, and Sundar Pichai of Alphabet and Google. The article places the accord against recent security incidents that caused OpenAI to halt training and pause releases. It does not explain how the new audits would be carried out.

[1]

Nothing new is legally required of the firms. Some of them, including Anthropic, Google, and OpenAI, had already committed to external audits, The Information reported. Many firms have also voluntarily joined highly secretive government safety tests. Trump called Tuesday’s agreement morally binding, even though Bloomberg reported that it carries no legal weight. He told reporters it was almost like a constitution: the biggest people in the world signed it, he signed it as president, and it was a form of protection. Near the end, the agreement says it may make sense to codify the recommendations. For now, Trump’s emphasis is that the country should trust the firms to regulate themselves.

[1]

Very few details have been released about how the new safety audits would work. Trump has posted only a summary of the accord on Truth Social. It is also unclear who would audit frontier models. Bloomberg reported that some of the firms poised to conduct those reviews have already met with the U.S. agency that runs voluntary pre-deployment tests, the Center for AI Standards and Innovation. Industry sources told Axios they were concerned that self-policing would not solve the safety problem now facing the industry.

The same article says Delaware and California are investigating whether governance failures contributed to recent loss-of-control incidents. That is a state-attorney-general inquiry, not a new duty created by the White House accord.

[1]

要点

  • About two dozen firms voluntarily committed to White House controls, including independent audits of internal controls, monitoring, and detection.
  • The named review topics are cybersecurity, biosecurity, chemical threats, and unintended model actions. The article does not describe audit steps.
  • Nothing new is legally required. Trump called the accord morally binding; Bloomberg said it has no legal weight. The text says codifying it later may make sense.
  • Few details explain how audits would work or who would run them. Trump has posted only a Truth Social summary.