Washington Wants 30-Day AI Model Review for Security

AI labs are now keeping their most powerful models behind vetted-access programs. This includes systems like Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber. These models can spot and exploit software flaws with great precision. For example, Project Glasswing, which used Mythos, helped Mozilla fix 271 vulnerabilities in Firefox 150. Cloudflare also found 2,000 bugs, with 400 being high or critical, using this technology. Overall, Mythos flagged over 10,000 critical or high-severity zero-days across about 50 partners, including major tech companies. It found flaws across operating systems and browsers, and even wrote working exploits. This rise in powerful AI capabilities has led to new directives from Washington. President Trump signed two AI directives this week, emphasizing national security considerations. The government is asking developers for up to 30 days of access to advanced AI models before they are released to other partners. This is a voluntary framework. The concern is that if such powerful AI falls into the wrong hands, it could weaponize flaws in critical infrastructure like power grids and banks. This period is seen as a crucial time to establish policies that will have significant long-term effects.