Nathan Lambert after the Black October wave: open weights for Claude Mythos were not the main risk, GLM-5.3 was

Nathan Lambert of AI2, in a follow-up piece, responds to the wave of security incidents he called Black October – it included critical vulnerabilities in PHP2Shell, GitLab, and a security incident at Binance. According to the Wall Street Journal, teams handling cyber incidents are stretched to the limit of their capacity. Even after this wave, Lambert maintains his earlier conclusion that if the weights of the Claude Mythos model had been released publicly as open weights, it would not have been a catastrophe, but rather an acceleration of an already existing risk trend. He identifies GLM-5.3 as the model that, in his view, truly crossed the dangerous capability threshold – more than a month has passed since its weights were released, and according to the author, there is no publicly available evidence of a fundamental change in the situation. As an example, he cites work by Epoch AI, which used the GLM-5.3 model to find an exploitable flaw in the Firefox browser allowing a user's files to be stolen merely by visiting a website – according to Lambert, however, this shows that causing real harm requires combining multiple steps. At the same time, he admits that he should concede more ground to part of the security community's position, and that a forced, costly overhaul of the approach to software security is inevitable.

In an earlier piece, which the current commentary builds on, Lambert criticized the debate over the cyber risks of open models as being split into two camps – proponents of banning open weights and proponents of keeping them for defense purposes – without any real discussion of what a ban would mean in practice or why Chinese companies consider their open releases to be safe. He described Anthropic's report on the cyber risks of the GLM-5.3 model as technically sound but not addressing these broader questions. According to Lambert, documented cyberattacks so far have been predominantly linked to closed models, and if open models were to be banned for security reasons, the same measures should logically apply to publicly accessible APIs of closed models as well. He regards open weights as a potentially key defense tool, especially for operation in air-gapped networks at sensitive institutions.

For details, see the source article.