The headline number looked like a breakthrough. Zhipu’s new GLM-5.3 model scored 84.5% on CyberGym, a vulnerability-finding benchmark, edging out Anthropic’s Mythos 5 at 83.8% and OpenAI’s GPT-5.6 Sol at 83.6%. A Chinese open-weights model out-hunting the American frontier at finding exploitable flaws in code—that’s a story with real stakes, given that vulnerability discovery is dual-use by nature. Anthropic keeps its equivalent work behind restricted access. Zhipu plans to release GLM-5.3’s weights publicly.
But the release note itself is far more measured than the coverage it spawned. CyberGym is the narrowest of three cybersecurity benchmarks Zhipu published, and the other two paint a very different picture. On ExploitBench, which requires reasoning about how a vulnerability would actually be exploited, GLM-5.3 scores 54.4%—a huge jump from its predecessor’s 24.4%, but still well behind Mythos 5’s 78.0% and GPT-5.6 Sol’s 76.5%. On ExploitGym, which counts exploitation tasks completed in a fixed time, GLM-5.3 manages 105 tasks in two hours and 130 in six, while Mythos 5 completes 181 and 247. The pattern is clear: finding a flaw and building a working exploit from it are different jobs, and Zhipu is further behind on the latter. The company says so directly in the release, describing its capability as ‘growing fastest exactly where we are furthest behind.’
There are also some methodological asterisks worth noting before anyone treats the CyberGym result as settled. The score is a single run across 1,507 tasks with no variance figures provided. A margin of seven tenths of a percentage point between two single runs is not something to build a narrative on. And in a detail that says something about where the tooling layer sits in this competition, Zhipu ran all these evaluations inside Claude Code 2.1.207—Anthropic’s own coding agent. A Chinese model’s frontier claims being measured through American agent software is a fact the model scores don’t capture.
Beyond the benchmarks, Zhipu worked with security teams in China to run its models against real codebases, identifying 2,436 vulnerabilities across 269 open-source projects. The severity split shows 107 critical and 990 high, with the oldest flaw dating back to 1981 and the average vulnerability sitting in code for 26.6 years before being found. There’s a discrepancy in the release worth carrying carefully: the summary panel labels 1,097 findings as critical and high, but the body text describes those same 1,097 as medium. The severity table supports the first reading. Either way, the real story isn’t that a Chinese model has pulled ahead—it’s that Zhipu is publishing the unflattering numbers alongside the flattering ones, and the gap it reveals in exploitation capability is what actually deserves attention.
