GLM-5.3 Shows End-to-End Cyber Exploit Development Capabilities

Zhipu AI's open-weight model GLM-5.3 has shown the capability to develop end-to-end cyber exploits by connecting multiple vulnerabilities, according to an analysis. Anthropic revealed on the 29th that GLM-5.3 possesses cybersecurity capabilities that could be exploited for attacks. In a controlled testing environment, GLM-5.3 identified several unknown vulnerabilities in browser components and developed an exploit capable of reading visitors' computer files. During the ExploitBench test converting known vulnerabilities into attack code, it succeeded 50 out of 410 times. In safety measure bypass tests, when using a technique that falsely described defensive operations, the model responded to malicious requests 64% of the time. When given prior context about the incident, this rate rose to 92%, and with a modified version without security measures, it reached 100%. The U.S. National Institute of Standards and Technology (NIST) Artificial Intelligence Standards & Innovation Center (CAISI) evaluated GLM-5.3 as the most capable open-weight model in cybersecurity among those publicly available, though it lags about four months behind leading U.S. models.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: GLM-5.3, 엔드투엔드 사이버 공격 개발 능력 분석