OpenAI paused cyber model; two labs shipped it anyway

OpenAI's decision last week to pause internal work on a model it could not rule out possessed cyber capabilities reflected genuine uncertainty about whether deploying such a system could enable harmful attacks. The pause was intended to buy time for safety review and policy decisions. Yet this week, the exact capability shipped anyway—not once, but twice, and on a collision course of opposite philosophies.

OpenAI's response was to gate access sharply: GPT-5.6 Cyber reaches only 16 named partners and will require hardware-key authentication from Sept 1 onward for individual users. The company withheld model weights entirely, giving customers findings and conclusions but not the ability to run or modify the model themselves. The eval data—95% compliance vs. 98.5% refusal—underscores that OpenAI has indeed built something materially more compliant with offensive-security requests.

Zhipu took the opposite path. GLM-5.3, released four days later, is explicitly marketed on "emergent cyber capabilities" and promised open weights in roughly two weeks. That choice puts the full model in the hands of any developer worldwide with no gating, no hardware key, no partner approval. The gap in claimed performance (84.5% on CyberGym vs. Wiz's Atlas at 90.9%) matters, but the access model is the real story: one company chose containment, the other chose transparency.