Nathan Lambert says the open-model cyber risk debate is missing the evidence

Nathan Lambert's argument in The Cyber Risk Discourse is Broken is that the fight over open-weight models has hardened into slogans while the evidence points the other way. He sorts the debate into three camps: lab leaders and national security people who treat open weights as an untenable risk, moderates who think a ban makes things worse, and Chinese firms that keep releasing strong open models based on their own risk judgments.

His strongest point is about where real attacks come from. People with classified briefings privately warn of a flood of attacks powered by Chinese open models, but Lambert notes that public records, including work like FelonyBench, mostly attribute actual cyberattacks to closed models, including OpenAI's. He uses Anthropic's own GLM-5.3 report to make the case: Claude's Mythos capability was first framed as a new class of cyber weapon, yet GLM-5.3 already crosses that bar, and more than a month after release there is little public sign of major harm. So a leaked open model would speed up existing risk rather than create a new kind of it.

The logic cuts both ways, he argues. If open weights deserve a ban on cyber grounds, public closed APIs would deserve one too, since their capabilities can outrun their guardrails. On air-gapped government networks, an open model may be the only usable defensive tool.

Why it matters

If you work on AI policy or model governance, this is a push to treat cyber risk as a trade-off with falsifiable predictions, not a settled case. Be skeptical of arguments you cannot check against public attack data, and watch whether the predicted wave of open-model harm actually shows up.