Open-weight artificial intelligence is moving closer to the performance of the industry's leading systems, according to a new evaluation of Z.ai's GLM-5.2 model. The report, produced by the AI safety nonprofit SaferAI, says the model is only a few months behind top-tier systems from OpenAI and Anthropic in cyber and biology-related capabilities.
What stands out most is the widening distance between raw capability and practical safeguards. SaferAI found that GLM-5.2 did not refuse the offensive cyber and dual-use biology tasks it was tested on, while some closed models apply stronger refusal layers and API-based controls.
That contrast highlights a central challenge for open-weight AI: once model weights are released, safety protections can be altered, removed, or bypassed on private hardware. In that environment, traditional guardrails such as classifiers, refusal training, and platform-level restrictions become much harder to enforce.
Experts note that even closed systems are not fully protected, since jailbreak methods can sometimes push models past their limits. But open-weight models raise a different question: how to preserve access to powerful tools while reducing the chance that harmful capabilities spread without oversight.
One proposed path is pre-training data filtering, which removes risky material before training begins. Researchers say this can help reduce hazardous biological knowledge without weakening overall performance, though it is less practical for cybersecurity because coding and offensive hacking skills often overlap.
As a result, developers are increasingly exploring selective restrictions, stronger pre-release testing, and clearer risk assessments. The broader debate is shifting from whether open-weight models can compete with frontier AI to how their release should be managed responsibly.
As open models continue to advance, the future may depend on whether the industry can make powerful AI broadly available while building safer standards around it.
