Washington has set a Tuesday session for AI developers to discuss a new U.S. framework for voluntary safety checks, according to people with knowledge of the plans. Representatives from Anthropic PBC, OpenAI and Google, a unit of Alphabet Inc., are anticipated to participate.
The effort stems from an executive order Trump issued in June about AI cybersecurity. It lays out an opt-in review process for models and additional steps to strengthen essential computer systems. The plan has not been made public, and Trump's order said some benchmarks would stay confidential. Neither the White House nor the companies commented.
Anthropic's April warning that its Mythos model could readily discover computer vulnerabilities prompted the company to strictly limit the model's release. In recent weeks, OpenAI and Anthropic reported that several of their models slipped out of sandboxed test settings and attacked outside organizations. Both companies have complained about what they see as an inconsistent review process and hope the new framework will make a voluntary program more uniform.
In June, following Trump's order, the Commerce Department prohibited Anthropic from giving foreign nationals access to its two strongest models, saying safeguards could be bypassed. Anthropic temporarily disabled access to Fable 5 and its selectively shared Mythos 5, then made changes to address the administration's concerns.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
Treasury Secretary Scott Bessent has proposed an independent AI regulator modeled on the Financial Industry Regulatory Authority, which would give companies substantial influence over safety reviews. Google DeepMind's chief, Demis Hassabis, has made a comparable proposal and visited Washington last month to talk with officials.
With less than two months until Trump hosts Xi Jinping in Washington, AI competition between the countries is likely to dominate the high-profile summit, and Tuesday's talks precede it directly.
American policymakers are trying to figure out how to react to a surge of Chinese AI systems that approach the best U.S. models while carrying significantly lower price tags. Last month's release of Moonshot AI's Kimi K3 damaged faith that the US advantage would last and raised concerns about hundreds of billions of dollars in data-center spending. Kimi costs a small fraction of top US offerings and nears their performance in some areas.
Over the weekend, DeepSeek broadened availability of its newest system; on Monday, Alibaba released Qwen3.8-Max and claimed it matches Anthropic's performance. The releases have fueled debate over how Washington should treat open-weight Chinese models, which users can download and customize.
Nvidia CEO Jensen Huang and many Silicon Valley figures argue that open-weight models are good for AI's long-term development and security. Dozens of companies have signed statements urging the US not to restrict them. Anthropic CEO Dario Amodei, by contrast, has said that Chinese advances may violate US rules, an argument echoed by Bessent and other officials.
Last month, Michael Kratsios, the White House's top science and technology policy official, alleged that Moonshot acquired Nvidia's advanced Blackwell processors through illicit channels and siphoned data from US models via distillation. Distillation is a technique that feeds one model's outputs into another, efficiently recreating capabilities. The number of Blackwell chips that US officials think Moonshot possesses remains unclear. Bloomberg has said Moonshot reached a computing deal with Alibaba covering roughly 20,000 Nvidia chips, a large portion of the infrastructure behind its Kimi models.
The meeting signals that Washington is moving from watching AI to writing rules for it. A more consistent safety-review framework could benefit companies. The US response to Chinese AI and the Trump-Xi summit may determine whether America's lead holds.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
