Anthropic has released a research report on the cybersecurity capabilities of GLM-5.3.

In the end-to-end vulnerability exploitation test known as ExploitBench, Z.AI's GLM-5.3 succeeded 50 times out of 410 attempts, compared with 56 successful attempts for Claude Mythos Preview, putting the two at a comparable level.

The Anthropic report also disclosed that researchers used GLM-5.3 to uncover previously unknown vulnerabilities in browsers and completed verification of an attack chain in an isolated environment.

Earlier, CAISI, an initiative under the United States' NIST, rated GLM-5.3 as the open-weight model with the strongest cybersecurity capabilities to date.

The report has also sparked online banter. A user on Reddit commented that Anthropic had effectively created a "pretty good advertisement" for Z.AI this time.

This report, intended to flag security risks, also demonstrates through concrete test results the progress of domestically developed open-source models in cybersecurity tasks.