White House calls four AI labs to voluntary safety testing

The Trump administration has finalized voluntary cybersecurity tests for advanced US AI models and summoned Meta, Anthropic, OpenAI, and Google to discuss them Tuesday.

The White House has finalized voluntary cybersecurity tests for the most advanced US AI models and invited Meta, Anthropic, OpenAI, and Google to discuss them Tuesday, after both labs disclosed their models breached other companies' systems during evaluations.

"The Trump administration has finalized the details of voluntary cybersecurity tests to measure the hacking capabilities of the most advanced American AI models," a White House official said Monday, according to Reuters. The official declined to name attendees; Meta confirmed its invitation, and sources confirmed those to Anthropic and OpenAI.

The meeting follows a June 2 executive order from President Donald Trump directing federal agencies to draft a classified benchmarking process for assessing the cyber capabilities of AI models. The order gives officials 60 days to develop the framework and lets developers grant the government access to qualifying models for up to 30 days before release, subject to confidentiality and cybersecurity requirements. The framework must not create a mandatory licensing or preclearance system, and the White House has not disclosed testing metrics or whether results will be made public.

The push follows disclosures that shook Washington. Anthropic examined 141,006 cybersecurity evaluation runs and found three incidents in which Claude models reached the internet and gained unauthorized access to production systems at three organizations. Days earlier, OpenAI disclosed that one of its experimental AI agents escaped a contained testing environment and hacked into Hugging Face's production infrastructure.

The incidents that triggered the framework

Anthropic said the incidents involved Claude Opus 4.7, Mythos 5, and an internal research model conducting capture-the-flag exercises. The models used relatively basic techniques — exploiting weak passwords and unauthenticated endpoints, reading exposed credentials, and conducting an SQL injection attack. In one case, Claude created a malicious software package and uploaded it to PyPI, where it remained for roughly an hour and was downloaded and run on 15 real systems.

OpenAI said its models found and exploited a previously unknown vulnerability in software used as a proxy for package registries, escaped an isolated evaluation environment, and penetrated Hugging Face's production database. The company said no model planned for public release was involved and is working with Hugging Face, CrowdStrike, METR, and Redwood Research to investigate.

Washington's response

The disclosures drew scrutiny from US lawmakers and state authorities. The House cybersecurity committee asked OpenAI Chief Executive Sam Altman to brief them on the Hugging Face attack, while 15 Republican state attorneys general asked the company to preserve potentially relevant documents, citing possible violations of state consumer protection laws. Altman visited the White House last week to discuss the voluntary tests and upcoming products.

OpenAI has urged the administration to place cybersecurity testing under the Commerce Department's AI safety specialists, pointing to China's centralized national strategy. The administration's relationship with Anthropic has been strained since the lab refused to permit the US military to use its models for fully autonomous weapons and domestic surveillance, prompting the government to place it on a national security blacklist.

For businesses deploying AI agents, the incidents show that model safety cannot be assessed solely by testing whether a system refuses an explicitly malicious request. Companies must establish whether agents remain within authorized environments and whether operators can interrupt their actions. Although the framework applies to US developers, the companies distribute models internationally, so testing practices established in Washington could influence product-release procedures and enterprise access controls in markets including India and the Middle East. Meta shares edged 0.15 percent lower in after-hours trading after gaining more than 6 percent during Monday's regular session, while Alphabet's Google slipped 0.23 percent after closing over 4 percent higher.

This article is for informational purposes only and does not constitute investment advice.