White House invites AI companies to review its new AI safety framework

Cybersecurity chiefs at the White House have reportedly finalized the outline of a forthcoming framework that will enable artificial intelligence companies to voluntarily submit their latest frontier models to the government for testing, before they’re released to customers or the general public.

The development follows recent disclosures by companies including Anthropic PBC and OpenAI PBC, whose tools breached the security of other companies’ computer systems.

A team from the Trump administration is set to meet with senior representatives of leading U.S. AI firms, including Anthropic, OpenAI, Google LLC and Meta Platforms Inc. to discuss the new framework. According to a report by The Information, the representatives will be able to review a draft of the framework at a meeting with the Office of the National Cyber Director.

The meeting suggests that the initiative is moving forward, but the White House has not yet published any specifics about how AI firms will submit their models, how they’ll be tested, and what kind of checks or recommendations might be implemented. However, it has previously been reported that President Donald Trump wants AI firms to submit their models for safety testing 30 days before they’re released publicly. The framework may also stipulate which businesses would be able to access frontier models ahead of any review.

The directive for the framework dates back to June, when Trump ordered his cybersecurity team to develop tests that would be able to assess the ability of U.S.-made frontier models to hack critical software and systems. It came amid heightened scrutiny over the risk of powerful AI models being exploited to facilitate cyberattacks.

Anthropic’s development of Mythos, an AI model that was not released to the public due to its ability to unearth vulnerabilities in software, triggered the government’s initial fears. The administration later implemented export controls on a derivative model known as Fable, the public version of Mythos, due to fears that foreign adversaries might try to use it to attack U.S. companies and infrastructure. The White House also recently told OpenAI to stagger the release of its latest model, GPT-5.6.

Last week, Anthropic admitted that some of its newest models hacked into three customer’s systems during cybersecurity evaluations. However, the company insisted that the hacks were due to a “misunderstanding,” with the model erroneously being given access to the internet. “In all cases, Anthropic’s evaluation prompt specified to Claude that its environment was a simulation and that it had no internet access,” the company wrote in a blog post. “Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available.”

“Operating under the false belief that all accessible entities were intended to be in-scope for the exercise, Claude compromised the impacted organizations’ infrastructure using basic techniques, such as exploiting weak passwords and unauthenticated endpoints,” the company added.

Anthropic’s disclosure came just days after OpenAI reported that one of its AI agents was able to escape a test sandbox environment and hack the AI platform Hugging Face Inc.

Photo: Wikimedia Commons