Z.ai and Concordia AI propose staged release rules for risky open-weight models

Zhipu’s Z.ai and Concordia AI, an independent AI-safety research organisation, jointly published a risk-management framework for open-weight models, arguing that because “a released model cannot be recalled,” risk decisions have to be made before publication rather than fixed afterwards. It sets out a six-stage process — identifying risks, setting thresholds across four dimensions (the deployment environment, who might misuse a model, what it enables, and how well society can absorb the resulting harm), then analysing, evaluating, mitigating and governing against them — and sorts models into green, yellow or red release zones. The hardest case, it argues, is the yellow zone: controlled-access measures that work for closed models, such as gating who can query a system, become unenforceable once weights are downloadable.

The report cites Z.ai’s own release of GLM-5.3 as a working example. After evaluations found the model unexpectedly capable of planning multi-stage cyber exploits, Z.ai gave vetted security partners and API access first and held back the public weights for about two weeks while it completed safety evaluations — a middle path between full release and none. The framework appeared a day before Anthropic reported that GLM-5.3’s own safeguards could be stripped out once its weights were public.