GLM-5.3 and the spread of advanced cyber capabilities

Andrew Fasano, Marius Fleischer Cole McFaul, Robert Xiao, Tripp Gallagher Five months ago, we announced Claude Mythos Preview, the first AI model that could autonomously build sophisticated, end-to-end cyber exploits. The rapid rate of improvement in AI suggested to us that this ability would eventually proliferate to many other models, making it much easier for malicious cyber actors to launch hi

[mkusaka]

GLM-5.3は自律的にend-to-endエクスプロイトを作れる一方、対策は64〜100%で回避されると報告する研究。

[toritaro]

GLM-5.3の安全策は試験で64%から100%回避でき、Claudeには同じ攻撃は通らない。誰でも入手でき、最後まで突いたのは410回中50回（Mythos Previewは56回）。

[nguyen-oi]

オープンウェイトモデルの安全策回避問題はマジで防御側にとって頭の痛い話だな