SaferAI: GLM-5.2 refused zero CyberGym offensive-security tasks that Claude Opus 4.7 refused consistently

TechCrunch·high signalXBlueskyLinkedInCopy link

A SaferAI report covered August 4 finds Z.ai's open-weight GLM-5.2 only a few months behind frontier models on cyber and dual-use biology capability, while shipping without a published safety framework, pre-deployment testing commitments, or risk assessment documentation. The sharpest data point: on the CyberGym benchmark GLM-5.2 refused none of the offensive cybersecurity tasks, whereas Claude Opus 4.7 refused so consistently that SaferAI could not complete the benchmark on it. Once weights are downloaded, API-level protections are unenforceable — the safeguards can simply be fine-tuned away.