White House talks with Anthropic shift to setting AI security rules

The White House is in discussions with Anthropic to develop a framework for evaluating the severity of security vulnerabilities in new AI models and determining when government intervention may be necessary. This initiative follows the recent export controls imposed on Anthropic, which led to the suspension of access to its latest AI models, Fable 5 and Mythos 5, due to concerns regarding a security flaw known in the industry as a jailbreak. Disagreements between administration officials and Anthropic's CEO, Dario Amodei, have arisen over the significance of this jailbreak. The rapid advancement of AI technology has outstripped the government's ability to effectively define and assess such security issues. The administration is striving to create a standardized method for evaluating incidents like this, highlighting its urgency in establishing regulations for powerful AI models that could pose risks to economic and national security. The negotiations are being led by Sarah Heck, Anthropic's head of public policy, and cofounder Tom Brown, with the aim of creating a common set of benchmarks for assessing future jailbreaks. These benchmarks will consider the extent of safeguards bypassed, the capabilities exposed, and the implications of any breaches. Although the export controls remain in place, the shift towards establishing technical standards suggests progress is being made in negotiations. Following a breakdown in talks when Anthropic refused to de-deploy Fable, citing limited vulnerability, the White House responded with export restrictions preventing foreign users from accessing the model. However, subsequent discussions over the weekend between senior administration officials and Anthropic leadership have facilitated in-person meetings aimed at reconciling differences. To address the situation, Anthropic sent senior researchers and security experts to the Commerce Department, indicating a commitment to work with the government on these critical issues. This collaboration underscores the recognition that no AI model can be entirely secure from hacking, and that clear rules are necessary to guide companies in assessing potential security risks.