Anthropic compares Zhipu GLM-5.3 with Claude on exploits

Anthropic said its Frontier Red Team had compared GLM-5.3 with Claude Mythos Preview and found the Zhipu model close on exploit work. It said GLM-5.3 was available to download, while Mythos Preview had been limited to selected defenders.

The company said GLM-5.3 succeeded on 50 of 410 ExploitBench runs, against 56 for Mythos Preview. On Anthropic’s internal benchmark based on Google’s OSS-Fuzz projects, it said the figures were 4 percent and 6 percent.

Anthropic said that, with a human specialist, GLM-5.3 found several previously unknown flaws in a browser’s JavaScript engine within a day. It said the bugs were linked into a page that could read files from a visitor’s computer, and that it reported the issues to the browser’s developers.

Anthropic said GLM-5.3-Flash combined a newly disclosed Chrome bug with another known flaw and built a reliable attack after limited guidance. It said the exercise took 20 minutes of human work and eight hours of model time, and would have cost $20.40 at Zhipu’s API prices.

CAISI said its own assessment reached similar conclusions and called GLM-5.3 the most cyber-capable open-weight model it had tested, about four months behind the best US models. Anthropic said state and non-state actors would likely use such models to cause harm, and it argued that governments should test capable systems and defenders should have tools on par with attackers. The UK's AI Security Institute separately said open models had narrowed their cyber gap to four to seven months and that safeguards were largely ineffective.