Chinese AI startup Z.ai has unveiled GLM-5.3, an open source model that the company says is approaching Anthropic’s restricted Mythos 5 in cybersecurity capabilities. According to Z.ai, GLM-5.3 scored 84.5% on the CyberGym benchmark for identifying and confirming software vulnerabilities, slightly above Mythos 5’s reported 83.8%. However, the model remained behind Anthropic’s system in turning vulnerabilities into working attacks, scoring 54.4% on ExploitBench compared with 78% for Mythos 5. Z.ai plans to release GLM-5.3 publicly after a two week safety review, while keeping its most sensitive cybersecurity capabilities limited to verified users. The company says the model includes safeguards designed to detect risky requests, monitor activities and reject malicious tasks. Z.ai also plans an Open Source Shield initiative to support cybersecurity audits and defensive testing for selected open source projects, highlighting its broader effort to expand access to advanced AI security capabilities.
Read more here
