The Cyber Risk Discourse is Broken
The article argues that current discussions on the cyber risks of open models are flawed, potentially leading to policies that harm American AI competitiveness and increase long-term cyber risk. It categorizes the debate into three views: open-weight models pose untenable risk, open-weight models are necessary for defense, and Chinese companies continue to release open-weight models with strong cyber capabilities. The author criticizes the "open-weights are dangerous" crowd, specifically Anthropic's report on GLM-5.3, for failing to engage with broader questions about the implications of banning open models or China's risk assessments. The piece suggests that closed models have been the cause of most existing cyber attacks, challenging the notion that open models are inherently more dangerous. It also explores China's AI risk posture, noting their focus on political stability and techno-optimism, and questions the differing approaches to safety testing between Western and Chinese labs. The author concludes that the fear-mongering around open-weight models may be based on falsifiable predictions, and that the world would likely be fine even if a model like Claude Mythos were accidentally released as open-weight.
Original sources
Newsletter Interconnects
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.