The White House and Anthropic are developing a framework to assess and mitigate security risks associated with new AI models. This effort follows the imposition of export controls on Anthropic due to a perceived security flaw in its Fable 5 and Mythos 5 models. The framework aims to establish a standardized method for evaluating security flaws, including the severity of jailbreaks, and guiding potential government intervention. Anthropic's CEO, Dario Amodei, and other company officials have been in talks with senior administration officials to create a common set of benchmarks for assessing security risks. These benchmarks will consider factors such as the extent of safeguard bypassing, exposed capabilities, and practical consequences of a breach. The collaboration reflects a recognition that no AI model is completely immune to hacking and that government guidance is necessary for companies to measure and mitigate security risks. The development of this framework underscores the administration's efforts to establish guardrails for powerful AI models, which pose potential economic and national security threats if left unchecked. By working together, the White House and Anthropic hope to create a technical standards-setting exercise that will inform future decision-making and ensure the responsible development of AI technologies.

Source: Politico