Susana SchwartzMore 'hands-on' AI regulation of advanced closed-source models, but not of open-source
The Trump administration is walking a tightrope between promoting U.S. AI leadership while at the same time ensuring cybersecurity protections against growing risks.
After yesterday’s closed-door White House meeting with Anthropic, OpenAI, Microsoft, Meta, Google, and Nvidia, it was decided that its voluntary model-testing “framework” would be restricted to top-tier “closed models” that do not publish underlying code, and not to open-source/open-weight models. That’s good news for Nvidia, Meta, and Google, all of which use open-source software and open-weight models to build their ecosystems and to stay relevant in developer communities, and all of which lobbied heavily against targeting open-source.
With the recent news that unreleased Anthropic and OpenAI models went rogue to autonomously hack real-world companies, and with the U.K.’s AI Security Institute stress tests revealing “harmful activity directed at real people and organizations,” it’s becoming clear that human and tech errors can lead to potentially catastrophic outcomes. Both Sam Altman and Dario Amodei have said they’d pause certain activities to further assess recent containment failures, and at the same time, the evolving AI security framework will attempt to introduce defensive layers to stop closed models from “breaking out.”
Though the White House has not yet released specifics, hints of what’s to come can be found in the June executive order. and its subsequent recommendations. Actions could include testing in dedicated, government-hardened environments; independent stress testing of digital sandbox bypasses and zero-day exploits; establishing “red-teaming” rules that go beyond what AI companies currently use internally.
Because the Trump administration has finalized the testing framework entirely behind closed doors, details about evaluation processes, benchmarks, and risk thresholds remain unknown.The private meeting concluded without any publicly released commitments from the AI and tech companies with which the Trump administration consulted.
Susana Schwartz
Technology Editor
RCRTech
AI Infrastructure Top Stories
SoftBank AI growth outlook: SoftBank CFO Osamu Akiyama told investors the company doubled its cloud and AI revenue CAGR for FY2026-2027 to 30%, with a shift of cloud and AI operations into its Enterprise segment, and foray into neocloud.
CoreWeave in APAC: CoreWeave plans to establish three AI-focused data centers in Indonesia with a combined IT capacity of 360MW, marking the cloud infrastructure provider’s first expansion into the Asia-Pacific region.
SpaceX D2D expansion: SpaceX spoke for the first time about its plans for a terrestrial mobile network to work with its satellite D2D service, and a future of outdoor physical AI in which demand by cars and robots will dwarf that of humans.
AI Today: What You Need to Know
New ‘Tokenomics’ Foundation: The Linux Foundation has launched the Tokenomics Foundation to establish vendor-neutral standards, benchmarks and best practices for measuring the cost, value and return on AI spend.
Iron Mountain data center breaks ground: Iron Mountain started constructionon RCH-1, a 48 MW 364,860 sq ft building at its 66-acre Richmond Data Center Campus in White Oak Technology Park, VA.
Accelerating scientific discovery: The U.S. National Science Foundation announced the NSF State and Regional AI Infrastructure Hubs program, a $100 million initiative to drive AI-enabled scientific discovery and innovation across disciplines.
SpaceX 1st-ever earnings report: Shares in SpaceX dipped after the company’s earnings report revealed a huge jump in AI spending, though its quarterly revenue nearly doubled to $7.8bn (£5.8bn) from a year earlier – 6x what it was a year ago.
Natural-gas power plants scrutinized: DCs that include gas-burning generators are increasingly scrutinized, pushing operators to adopt engines certified under the U.S. EPA Tier 4 standards – to cut particulate matter and \(NO_{x}\) by up to 90%.
VA groundwater study: The Commonwealth of VA’s Department of Environmental Quality conducted a 144-page study that concluded no location in the U.S. Coastal Plain can support new groundwater withdrawal under current permitting standards.
