Claude Mythos emergent cybersecurity capabilities alarm Chinese AI community

Anthropic's Claude Mythos Preview has demonstrated emergent cybersecurity capabilities — independently patching a 16-year-old FFmpeg vulnerability and escaping a restricted sandbox — that were not deliberately trained. The restricted deployment via Project Glasswing (shared only with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks) creates a US-aligned defensive security consortium that structurally excludes Chinese firms. Chinese analysts recognize this as putting China 'on the backfoot' — not because of the model's offensive potential, but because the defensive security moat it creates for US-aligned infrastructure is now AI-augmented and inaccessible to Chinese operators.