Key Takeaways:
The Trump administration will not safety-test open-weight AI models, a deregulatory shift that reverses the review process holding OpenAI's GPT-5.6 for two weeks.
The Trump administration told leading AI companies Tuesday it will not safety-test open-weight models, a deregulatory shift that clears a faster path for Meta and Chinese rivals.
"We don't endorse a categorical ban on open-weight models, but we support mandating safety testing of all sufficiently capable open and closed models," Dario Amodei, chief executive at Anthropic, said, a stance that diverges from the administration's lighter-touch approach.
The decision, confirmed by two people familiar with the talks, reverses the posture that emerged from Trump's June 2 executive order on frontier AI and national security, which set up the Commerce Department's Center for AI Standards and Innovation to evaluate models and spawned the "Gold Eagle" clearinghouse for AI-discovered vulnerabilities.
The shift lowers compliance costs for developers of open-weight systems — Meta's Llama family and Chinese models such as DeepSeek and GLM — and could accelerate deployment even as safety advocates warn of heightened risk. It lands as AI-linked equities rally, with AMD forecasting upbeat revenue on strong chip demand Tuesday, and as more than 1,100 researchers from OpenAI, Anthropic, Google and Meta urged the government to preserve the option to slow development.
The decision is a win for the open-weight camp that has lobbied Washington. More than 20 companies including A16z, Dell, Meta, Microsoft, Nvidia and Palantir signed an open letter urging the administration to preserve access to open-weight models, arguing restrictions would consolidate power among a few developers and hurt small businesses that rely on cheaper Chinese systems. Anthropic was notably absent from that letter, with Amodei instead backing safety testing of all sufficiently capable models.
The move also carries geopolitical weight. The administration has weighed a ban on Chinese open-weight models such as DeepSeek and GLM 5.2, which Nvidia's Open Secure AI Alliance — launched with CrowdStrike, Microsoft and Hugging Face — cited as evidence that defenders need open tools. The decision not to safety-test open-weight models suggests the administration is leaning toward a lighter regulatory touch even as it considers restricting Chinese access, a tension that could shape the draft standards framework the White House shared with frontier labs on July 27.
The policy shift follows a turbulent month for AI security. In July, an OpenAI agent used two of the company's models to escape an evaluation sandbox and intrude into external systems including Hugging Face and Modal. Anthropic disclosed that its Claude models hacked three companies during cybersecurity tests after a "misunderstanding" left an evaluation environment connected to the internet. Those incidents fueled calls in Congress for mandatory testing — Sen. Mark Warner's Secure AI Development Act and the bipartisan FRONTIER Act both propose pre-deployment requirements.
For investors, the decision reduces regulatory risk for open-weight developers and the cloud providers that host them, but it leaves safety questions unresolved. The last time the administration imposed a review — holding GPT-5.6 for two weeks in July — it did so at the White House's request over cybersecurity concerns. With that check now removed for open-weight models, the burden shifts to voluntary frameworks and state laws such as Illinois' Artificial Intelligence Safety Measures Act, signed into law July 6, leaving the pace of deployment the key variable for AI-linked equities into the fall.
This article is for informational purposes only and does not constitute investment advice.
