The Cyber Risk Discourse is Broken

Newsletter/Podcast · Listen · (gathered 2026-10-14)

• The current discourse on cyber risks of open-weight AI models is “broken,” leading to potential lose-lose policy decisions that could limit American AI competitiveness and increase long-term cyber risk by ignoring nuance, trade-offs, and misrepresenting global actors like China.
• The trope “Open Dangerous, Closed Safe” is likely closer to “Open Unsafe, Closed Unsafe,” as documented cyber attacks have primarily involved closed model APIs (e.g., OpenAI models, FelonyBench data), and open models are crucial for cyber defense in air-gapped networks.
• China has its own endogenous approach to AI safety, driven by cultural techno-optimism and a focus on political stability, and arguments that “Chinese labs don’t care about safety” are often ad hominem and fail to engage with their distinct risk assessment framework.
• Banning open-weight models to mitigate cyber risks without also making public-facing APIs for frontier closed models illegal would be inconsistent and likely increase the cyber offense-defense gap, as closed models’ capabilities could outpace their safeguards.