Z.ai and Concordia AI propose six stages for managing open-weight AI risk

Zhipu’s Hong Kong Office

Credit: Zhipu

Chinese AI company Z.ai and Beijing safety consultancy Concordia AI have published a framework for managing the risks of open-weight AI models. Anyone can download, modify and run the trained parameters of such models.

The two released the report on Monday, Chong Ming Lee reported for the South China Morning Post.

The report “provides the first comprehensive, evidence-based foundation for navigating these tensions” between openness and safety, its authors write.

Why open weights are different

The framework starts from three problems. A release is difficult or impossible to reverse. Fine-tuning can strip out safeguards. And once weights are out, developers can no longer monitor how people use the model.

“Safety training can be undone with a small number of harmful training examples,” the report says.

So the report focuses on measures that act before release, or that keep working after the weights are out. It lists training data curation early in development, and staged release instead of a single open-or-closed choice. Z.ai’s GLM-5.3 is its example: vetted security partners tested the model first, and the full weights came out only after safety evaluations.

The framework sets out six stages, from identifying risks to governing them:

Risk identification covers misuse and accidents. Risk thresholds define unacceptable risk along four dimensions. They are where the model runs, who might misuse it, what it enables and how well society can absorb the harm. Risk analysis runs before development, before deployment and after it.

Risk evaluation sorts models into green, yellow and red zones: full release, restricted or staged release, or suspension. Risk mitigation applies protections across the model’s life, and risk governance sets out oversight and accountability.

The report is a companion to the Frontier AI Risk Management Framework 2.0 from Shanghai AI Lab and Concordia AI, published in July.

Context

Z.ai raised $5bn in Hong Kong this month. Last week it apologised over its ZCode coding tool and open-sourced it. In August, a White House technology strategy left open-weight AI off its list of critical technologies.

Get the TNW newsletter

Get the most important tech news in your inbox each week.