Anthropic assessed the cyber capabilities of GLM-5.3

On September 29, 2026, Anthropic published a study titled “GLM-5.3 and the spread of advanced cyber capabilities,” which found that Z.ai’s open model GLM-5.3 nearly matched Claude Mythos Preview for the first time among open-weight models in autonomous cyber-exploit creation, and in manual sessions it independently discovered unknown 0-day vulnerabilities in a browser JS engine.

Compiled by Sergey KostenchukPublished 2026-09-30Updated 2026-09-30

2026-09-30 Safety Anthropic

🛡 Anthropic assessed the cyber capabilities of GLM-5.3

Z.ai’s open model GLM-5.3 nearly caught up to the closed Claude Mythos Preview for the first time in creating cyber-exploits: 50 out of 410 in ExploitBench versus 56. In manual tests, it independently discovered unknown 0-days in a browser JS engine.

🌍 Open weights are publicly available, while top US models are only accessible through verified access. The “obliterated” version of GLM-5.3 without built-in refusals ($1,200–4,400) bypasses protections in 100% of cases. NIST CAISI: the lag of the open model behind the US frontier is about four months.

👤 A working exploit chain for Chrome CVE-2026-11645 cost 20 minutes of human attention and about $20.40 at API prices. Update your browsers and operating systems as soon as possible, and do not consider the refusals of open models as a protection.

Source 1: https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities Source 2: https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities

Sources