Anthropic Report: Zhipu GLM-5.3 Can Autonomously Build End-to-End Exploits

Anthropic released a report on September 29 stating that Zhipu's GLM-5.3 demonstrates strong cybersecurity capabilities. In ExploitBench end-to-end vulnerability exploitation tests, GLM-5.3 completed 50 out of 410 attempts, close to the 56 achieved by Anthropic's Claude Mythos Preview.

The report also said researchers used GLM-5.3 to discover unknown vulnerabilities in browsers and verified attack chains in an isolated environment. Anthropic noted that, like Claude Mythos Preview, GLM-5.3 can autonomously construct end-to-end network attack exploits. However, its guardrails were found to be relatively weak: in simulated tests, attackers could bypass them 64% to 100% of the time using simple techniques, while similar attacks did not succeed against protected Claude models.

GLM-5.3, released by Zhipu on August 14, 2026, uses an open-weight model and includes a "defense-in-depth" risk review system. Zhipu said at the time that the model's own safety alignment remains effective after open-weight release. The download was delayed by two weeks for security assessment and hardening.

According to Zhipu's disclosure on September 16, cybersecurity became the first vertical scenario to achieve scaled revenue one month after GLM-5.3's release, with 100 security companies integrating GLM models into their products and business processes, covering major domestic security vendors, internet firms, research institutions, and financial organizations.