China’s Zhipu AI claims new GLM-5.2 closes gap with Anthropic on security tasks

Z.ai released GLM-5.2 on June 28, 2026, and is pitching it as an open-weight model built for long-horizon work, with a usable 1M-token context for project-scale engineering. The company says it is the strongest open-source model on standard coding benchmarks, citing scores of 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, up from 63.5 and 58.4 on GLM-5.1. The sharper question is not whether the numbers improved, but what it actually means to match Anthropic’s Claude Mythos in cybersecurity, where bug-finding, triage and disclosure are not the same thing.

Anthropic has described Claude Mythos as strikingly capable at computer security tasks and has framed it as part of a broader security push. On April 7, 2026, the company launched Project Glasswing with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Anthropic has said the bottleneck in modern vulnerability work is shifting away from finding flaws and toward triaging them, disclosing them responsibly and patching them fast enough to keep users safe.

That matters because the benchmark story is only part of the verification problem. If GLM-5.2 is being measured against Mythos in narrow bug-finding settings, the comparison says more about a specific workflow than about broad frontier AI parity. Z.ai’s own framing puts the model in coding and agentic engineering, not general intelligence, and other public assessments have said the gap with Anthropic and OpenAI remains wider on general tasks even as it narrows on finding bugs.

The geopolitical stakes are rising quickly. The United States has allowed Anthropic to release Mythos to some trusted partners after national-security concerns led to tighter controls, a sign that access to advanced security models is now being handled like a strategic capability rather than a routine product launch. China is moving in the same direction. On June 24, 2026, 360 Security Technology unveiled Mythos-like cyber tools, underscoring that Chinese firms are openly chasing parity in AI-assisted vulnerability discovery.

Z.ai’s open-weight approach changes the balance in both directions. It can make a security model easier for defenders, researchers and bug hunters to run, inspect and adapt. It can also make the same capabilities easier to repurpose for offensive use. The immediate contest is not over abstract AI supremacy. It is over who can find, verify and fix software weaknesses fastest, and whether an open model lowers the barrier for the people trying to do both.