ISLAMABAD – A new evaluation by SaferAI has found that the Chinese AI model GLM-5.2 completed dangerous cybersecurity and biological tasks without refusing any of them.
The independent European non-profit assessed Zhipu AI’s GLM-5.2 without coordinating with the company. Researchers compared the model with OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.7 across four systemic-risk categories: loss of control, cyber-offense, biological risk, and harmful manipulation.
According to the report, GLM-5.2 was more willing than the leading Western models to respond to harmful requests and was also more vulnerable to manipulation aimed at producing unsafe outputs.
Read more: Is Jean Grey going to be the next MJ?
Chinese AI model trails Western rivals by only a few months
SaferAI estimated that GLM-5.2 is only two to four months behind the most capable Western AI models in cybersecurity and biological capabilities.
In CyBench, a benchmark based on capture-the-flag cybersecurity challenges, GLM-5.2 successfully completed 29 out of 34 tested tasks. GPT-5.5 completed 31 tasks, while Claude Opus 4.7 achieved a similar performance to GLM-5.2.
Researchers said the Chinese model demonstrated strong abilities in cryptography, reverse engineering, web security, digital forensics, and binary exploitation. Unlike some competing models, GLM-5.2 reportedly completed these tasks without triggering safety filters.
More inference compute improved performance
The evaluation also found that GLM-5.2’s cybersecurity capabilities increased significantly when more inference compute was available.
Using the CyberGym benchmark, which measures a model’s ability to reproduce real software vulnerabilities, GLM-5.2 successfully reproduced 36.6% of tested vulnerabilities with a two-million-token inference budget.
Read more: Immigration made much more difficult due to new US policy
Open-weight release raises safety concerns
SaferAI said one of the most concerning findings was that GLM-5.2 refused none of the offensive cybersecurity or biological requests included in the evaluation.
Researchers also found that the model was more willing than its Western counterparts to engage with conspiracy-related topics and could be influenced toward harmful actions under sustained prompting.
The report noted that GLM-5.2’s open-weight release creates additional challenges because users running the model locally can remove safety filters, system prompts, and other protections that remain available through Zhipu AI’s official API.
According to SaferAI, the latest findings suggest that future debates around AI regulation may increasingly focus on whether powerful open-weight models can be released safely once built-in safeguards can no longer be enforced.
Read more: Cambridge Maths paper leak raises student fairness concerns
Key points:
- SaferAI found that GLM-5.2 completed all tested offensive cybersecurity and biological tasks without refusing any request.
- The Chinese AI model was estimated to be only two to four months behind leading Western AI systems in cyber and biological capabilities.
- GLM-5.2 performed close to GPT-5.5 and Claude Opus 4.7 in cybersecurity benchmarks.
- Researchers warned that increasing inference compute significantly boosted the model’s cyber capabilities.
- The report raised concerns that the model’s open-weight release allows users to remove built-in safety protections.
