The new AI model that's alarming Washington | The Economist

TL;DR

Anthropic's secret AI model Mythos demonstrates unprecedented hacking capability, alarming U.S. officials and prompting a potential policy shift in the hands-off Trump administration.

Key Points

• 1.Mythos represents a dramatic leap in AI-enabled cyberattack capability. Unlike Opus 4.6 which succeeded twice in hundreds of hacking attempts, Mythos achieved 181 successful exploits and 27 full attack chains reaching the Windows registry.
• 2.Anthropic restricted Mythos to only 11 handpicked partners including Apple, Microsoft, and the Linux Foundation. The strategy mirrors OpenAI's 2019 GPT-2 delay, hoping trusted companies can patch vulnerabilities before a similar model reaches the public.
• 3.Washington officials, including Treasury Secretary Scott Bessant and Fed Chair Jay Powell, held emergency bank meetings within a week of the announcement. Even Trump's typically hands-off, accelerationist administration has been visibly alarmed by Mythos's capabilities.
• 4.The likely U.S. government response will be informal and industry-led rather than heavy regulatory intervention. Officials are expected to mandate limited staged releases and develop a certification framework co-designed by major AI companies and government.
• 5.The upcoming Trump-Xi summit in May may open early dialogue on AI safety coordination. Despite massive mistrust, the threat of critical infrastructure attacks by rogue actors is seen as a shared risk that could motivate both powers to discuss standards.