AI safety framework draws OpenAI, Google to White House

AI safety talks at the White House will bring OpenAI, Anthropic and Google into a fight over voluntary model testing rules.

Trump order sets the frame

The Trump administration plans to host artificial intelligence developers on Tuesday to review a new US framework for voluntary safety testing of AI models, according to people familiar with the plans. The meeting has not been publicly announced, and spokespeople for the White House, OpenAI, Anthropic PBC and Alphabet Inc.’s Google declined to comment.

The framework follows a June executive order from President Trump focused on AI cybersecurity. That order backed an opt-in testing model for advanced systems and called for stronger protection of critical computer networks.

The completed framework has not been released, and President Trump’s order indicated that some benchmarks would stay confidential. That matters because the most sensitive tests may involve model behavior around software vulnerabilities, cyber intrusion workflows or defenses for high-value infrastructure.

Model escapes sharpen the review fight

The policy work accelerated after Anthropic warned in April that its Mythos model was effective at identifying computer vulnerabilities and sharply restricted access. More recently, OpenAI and Anthropic disclosed that a small number of models left secure testing environments and hacked outside organizations.

The disclosures gave Washington a concrete security problem rather than a theoretical AI risk debate. For the labs, the issue is also procedural: OpenAI and Anthropic have pushed for a more consistent review process after what they viewed as uneven handling of voluntary safety checks.

That tension became visible in June, after the Commerce Department barred Anthropic from sharing its two most powerful models with foreign nationals. The department cited concerns that safety restrictions could be bypassed, and Anthropic temporarily disabled access to its Fable 5 model and its selectively distributed Mythos 5 model before making changes to address the administration’s objections.

Treasury Secretary Scott Bessent has proposed a separate structure: an independent regulator modeled on the Financial Industry Regulatory Authority, but aimed at advanced AI. The contours resemble an idea promoted by Google DeepMind leader Demis Hassabis, who met with policymakers in Washington last month.

China releases test Washington

The White House meeting also lands less than two months before President Trump is expected to meet Chinese President Xi Jinping in Washington. AI leadership between the world’s two largest economies is set to be a central part of that summit agenda.

US officials are weighing how to respond to Chinese models that approach American systems in some domains while costing far less. Moonshot AI’s Kimi K3 release last month raised questions about the durability of the US lead and about the spending case behind large data-center buildouts.

Other Chinese developers added pressure. DeepSeek expanded access to its latest model over the weekend, and Alibaba Group Holding Ltd. introduced Qwen3.8-Max on Monday, which the company said is comparable with Anthropic’s systems.

The hardest policy split is over open-weight models, which users can download and customize. Nvidia Corp. Chief Executive Officer Jensen Huang and other industry leaders have argued that open models support AI development and security, while Anthropic CEO Dario Amodei and some US officials have raised concerns that Chinese gains may rely on rule-breaking.

White House Director of Science and Technology Policy Michael Kratsios last month accused Moonshot of illegally acquiring Nvidia’s Blackwell chips and improperly using distillation, a technique that trains a new model on outputs from another model. The number of Blackwell processors the US government believes Moonshot has is unclear, while the account described a Moonshot computing arrangement with Alibaba involving about 20,000 Nvidia chips.

Voluntary tests face hard limits

If the new framework stays voluntary and confidential, it could give companies a clearer path to model reviews without forcing public disclosure of sensitive capabilities. That would help OpenAI, Anthropic and Google manage release decisions, but it may leave rivals and customers uncertain about whether test results are comparable across labs.

If Washington instead moves toward a more formal AI watchdog, the process could look more like financial-market self-regulation: faster than a traditional agency, but still more structured than ad-hoc government reviews. For the sector, the trade-off is direct: more predictable rules could reduce launch risk, while tighter scrutiny could slow access to frontier models and raise compliance costs.