Claude Mythos AI model fuels U.S.-China race on cyberattack capabilities

Anthropic previewed the Claude Mythos model in April 2026 but held back a public release, citing safety concerns over its advanced cyberattack abilities. In a controlled-access program called Project Glasswing, the model was credited with discovering more than 10,000 high‑severity software flaws, including a decades‑old remote code execution bug, demonstrating a capacity to autonomously locate and exploit vulnerabilities faster than human teams.

China has announced plans to field a comparable model by the end of 2026, with a Zhipu AI co‑founder publicly stating the target and a U.S. think‑tank projecting an early‑2027 timeline. Beijing’s approach contrasts with the United States, which spent months in 2026 devising an AI governance framework after the Glasswing rollout. Since 2021, China has built a mandatory registration and compliance system for generative AI, processing 748 service registrations by the close of 2025, positioning the country to integrate such advanced models within an established regulatory regime.