Argues that restricting open-weight models over cyber risk would not stop attackers and would cut US competitiveness and defensive capability. Engineers relying on open models should track where this policy debate is heading.
Key takeaways · AI-distilled
Lambert notes that public evidence so far documents closed models as the cause of most existing AI-assisted cyber attacks, which cuts against claims from classified briefings that Chinese open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models are driving an onslaught.
He argues that if open-weight models must be banned to slow cyber risk, consistency would also require outlawing public APIs to frontier closed models, whose safeguards are stronger but far from perfect.
Lambert warns a ban on open weights while closed models advance would widen the offense-defense gap, since some defenders, such as sensitive government agencies on air-gapped networks, can only deploy open-weight models.
On China, he says labs register major releases and evaluations with the government, and he frames the real question as the minimum compute a lab should spend on safety testing, noting full evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → of a model like Kimi K3 could cost tens of millions.
Lambert argues GLM-5.3 crossed the capability threshold once attributed to Claude Mythos with little public evidence of harm a month after release, and calls the forecast of crippling open-weight cyber harm a falsifiable prediction he expects to fail.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
