Nathan Lambert argues cyber risk discourse on open-weight models is broken due to ignored trade-offs

Nathan Lambert argues that the current debate over the cyber risks of open-weight AI models is flawed because it ignores complex trade-offs and ecosystem dynamics. He posits that policy decisions driven by fear of open models may simultaneously limit American AI competitiveness and increase long-term cyber risk.

The Flaws in Anti-Open-Weight Arguments

Lambert identifies three main perspectives in the debate: frontier labs and national security officials viewing open weights as an untenable risk; moderates arguing open weights are necessary for defense; and Chinese companies releasing models based on their own risk assessments. He criticizes the "anti open-weight alliance" for failing to engage with cross-cutting questions, such as why Chinese companies deem their models safe to release.

The author challenges the narrative that open models are uniquely dangerous, noting that public information shows closed models have been documented as the cause of most existing cyber attacks. He suggests that the trope "Open Dangerous, Closed Safe" may actually be "Open Unsafe, Closed Unsafe," as closed model APIs are highly accessible despite safeguards. Lambert argues that if open-weight models are banned to slow cyber risk diffusion, public-facing APIs for frontier closed models should also be restricted, as the latter's capabilities may outpace their guardrails.

China’s AI Risk Posture

Lambert explains that Chinese companies approach AI safety through an endogenous lens influenced by techno-optimism and a focus on political stability. Unlike the U.S., where safety debates are prominent, Chinese labs register model releases with the government, though the extent of cyber-specific evaluation frameworks remains unclear. He notes that personal risks for Chinese researchers are higher, and labs are incentivized to compete rather than spend tens of millions on comprehensive safety evaluations that might delay releases.

Evidence Against Catastrophic Predictions

The article discusses the hype surrounding Claude Mythos, which was previewed as a potential cyber weapon but did not result in mass societal destabilization. Lambert points to GLM-5.3 as the model that actually crossed capability thresholds, yet over a month after its release, there is little public evidence of unprecedented harm. He concludes that proponents of open-weight fear-mongering are making falsifiable predictions that appear to be wrong, suggesting that open-weight models may actually serve as essential tools for cyber defense in air-gapped environments.