Z.ai's GLM-5.2 Refused Zero Attacks That Claude Opus Refused

GLM-5.2 refused zero of the offensive tasks that stalled Claude Opus 4.7's CyberGym benchmark, per a SaferAI report TechCrunch covered August 4.

Z.ai's open-weight model trails the frontier by only months on cyber and dual-use biology capability, per SaferAI. It shipped with no published safety framework, no pre-deployment testing commitment, and no risk documentation.

SaferAI couldn't score Claude Opus 4.7 on CyberGym at all, since it declined every offensive task in the set. GLM-5.2 completed the same set without a single refusal, per the report.

Yes, but the report doesn't say what GLM-5.2 accomplished once it stopped refusing, only that it never refused. Attempting a task and succeeding at one aren't the same measurement, and TechCrunch's coverage doesn't close that gap.

It doesn't need to close it for the core problem to hold. Once weights are downloaded, there's no API left to gate. Whatever GLM-5.2 will or won't do, it does locally, outside Z.ai's reach, per the report.

That zero-refusal score is GLM-5.2's stock behavior, not a jailbreak, and nobody can patch it once it's out.