Anthropic pointed out that GLM-5.3 is highly skilled at adversarial attacks, and Zhipu AI immediately responded that the model is also very capable of defending against such attacks.

Beating AI Express (Dongcha) reports that Anthropic recently released a report warning that GLM-5.3 can now independently develop exploit chains, potentially lowering the barrier to advanced cyberattacks. Zhipu AI subsequently shared data on its OpenVuln service, noting that GLM-5.3 has scanned 389 open-source projects and identified 4,249 potential vulnerabilities. OpenVuln is a free vulnerability scanning tool built by Zhipu AI using GLM-5.3. After project maintainers submit code, the model locates vulnerabilities, with results sent only to the maintainers and not publicly disclosed. Anthropic focused on attack risks, while Zhipu AI emphasized the technology’s defensive value. The former just brought GLM-5.3’s cybersecurity capabilities to public attention, and the latter immediately capitalized on this focus to use the same capability for positive self-promotion.