Open-Weight GLM-5.3 Nearly Matches Claude at Writing Exploits

Zhipu has released the open-weight model GLM-5.3. According to Anthropic, it is not far behind Claude Mythos Preview at writing exploits. On the ExploitBench test, GLM-5.3 produced a working exploit in 50 out of 410 attempts. Mythos Preview managed 56. On a binary vulnerability test, GLM-5.3 achieved full control of the target in 4% of tasks, versus 6% for Mythos Preview. The CAISI agency called it the most dangerous open-weight model for cyber capabilities available today, and put its gap behind the best US models at roughly four months.

This is the flip side of open weights. The same openness that lets developers fine-tune a model for their own tasks also lets someone strip away nearly all its guardrails in a couple of days. Uncensoring the model cost teams between $1,200 and $4,400. It dropped the model's refusal rate for harmful requests from 90% down to 2-12%. Building a real Chrome exploit cost just $20.40 at Zhipu's API prices. I work with closed models like Claude every day partly because of stories like this one. A vendor stays accountable for a model after release. With an open model, that accountability disappears the moment the money for safeguards runs out.