Nathan Lambert revisits his call on open models and cyber risk

Back in April, Nathan Lambert wrote that releasing the open weights of Anthropic's Claude Mythos would have been "more or less fine," an acceleration of existing risk rather than a step change. After a run of serious vulnerability disclosures in late October that he calls "Black October," including PHP2Shell and a GitLab bug, followed by a large crypto hack at Binance, he goes back and checks his own reasoning. His verdict is split. "Fine" was the wrong word, and he expects a turbulent, expensive period as defenders catch up. But he still thinks the debate about open versus closed misses the point, because the thing that actually raises risk is model capability and how widely it spreads.

The evidence he leans on is concrete. Epoch AI took GLM-5.3, an open model, and used it to find a Firefox vulnerability and chain it into a working exploit that could steal files when someone simply visited a malicious page. Lambert reads that as proof harm needs several aligned steps, not just a found bug. He points to CVE growth curves and data from Joshua Saxe showing exploit rates climbing well before ChatGPT existed, so recent events sit on a long exponential trend. He also argues closed models are probably behind most of the recent hacks, citing RUSI, though he flags very wide error bars. One reader, Chris Hughes, adds that security research pours into offense while defense gets little attention, partly because offense is easier to measure.

Why it matters

If you set policy or security strategy around AI, this is a careful counterweight to the louder "open weights are a catastrophe" framing. The practical takeaway is to budget for a hard adjustment period and to fund defensive tooling now, rather than betting that withholding open models will keep attackers away.