The White House is discussing a model security assessment framework with AI company Anthropic, aiming to establish a unified method for classifying security flaws in next-generation AI models and determining whether government intervention is necessary. This comes after the U.S. imposed export restrictions on Anthropic's latest model due to a security issue deemed akin to "jailbreaking."

Establish unified standards around "jailbreaking" vulnerabilities

According to reports, this framework will be used to assess the severity of similar future incidents, focusing on three key aspects: the extent to which safeguards are bypassed, the capabilities of the exposed models, and the real-world consequences of the vulnerabilities.

Currently, significant differences remain between governments and businesses on these issues. Previously, Anthropic CEO Dario Amodei and government officials disagreed on whether the vulnerabilities constituted a major security problem. The report points out that AI technology is advancing too rapidly, and existing government systems are insufficient to provide a consistent assessment of such controversies.

Negotiations continue after export restrictions

The White House had previously imposed export restrictions on Anthropic, prohibiting overseas users from accessing its latest models, the Fable 5 and Mythos 5. The company subsequently suspended external services for these two models.

The report noted that negotiations between the two sides nearly broke down last Friday. This was because Anthropic refused to remove Fable as required by the government, arguing that the vulnerability had limited impact and was insufficient to be considered a serious security flaw. Subsequently, the White House imposed export restrictions, forcing the company to withdraw the relevant models from the market.

However, communication between the two sides resumed over the weekend. U.S. Commerce Secretary Howard Lutnick, National Cyber Director Sean Cairncross, and Anthropic co-founder Tom Brown participated in several lengthy calls. Following this, the two sides held nearly a week of in-person talks in Washington.

The White House is accelerating the implementation of AI safety rules.

Reports indicate that Anthropic's representatives involved in the negotiations included Sarah Heck, head of public policy, and co-founder Tom Brown. The company also sent senior researchers and security experts to the U.S. Department of Commerce on Monday to continue discussions with government officials.

This round of discussion also reflects a more realistic assessment: no AI model is completely immune to hacking. Therefore, the government wants to first clarify the standards by which companies assess security risks before deciding when to take restrictive measures.

This direction echoes the discussions at the recent G7 summit. Several leading AI companies and some national leaders have mentioned the need to quickly develop clearer standards for measuring model security in order to address the potential economic and national security risks posed by increasingly capable AI systems.