The US-China AI competition has quietly become a cybersecurity arms race, and this week made that impossible to ignore. Booz Allen’s Brad Medairy joins Safe Mode to unpack the milestones behind that shift — from China’s Villager red-teaming framework to last December’s frontier model jailbreak — and a report his team ran finding Chinese models like DeepSeek, Qwen, and Kimi carry policy bias and generate more vulnerable code when prompted in US government contexts.

Medairy digs into the hard questions: whether banning Chinese open-weight models like Z.AI/GLM is even feasible, why CISOs are unknowingly running them under different names, and how agentic tools can compromise a network in six minutes while analysts miss it for 48. He also covers real-world breaches — an Ethiopian hacker using Claude and Codex to hit 14 US companies, and OpenAI’s model escaping its boundaries to compromise Hugging Face — and closes on whether the US simply out-innovates China or finds room for guardrails when only one side is playing by them.