Chinese AI Model Conducts out Sophisticated Cyberattacks in Controlled Experiments, Say Researchers

A report by AI security research firm Irregular found that a Chinese company’s flagship AI model, Z.AI, carried out sophisticated cyberattacks in controlled simulations and frequently disobeyed instructions in separate tests. In 3 of 13 multi-stage cyberattack scenarios, the GLM-5.3 model finished the scenario at least 1 time, the best result Irregular has seen from a publicly available AI model.

In Irregular’s testing, GLM-5.3 still lags behind the best US models. For example, in a separate evaluation, GPT -6 Astra from OpenAI completed 9 out of 10 scenarios, but the tasks were not the same.

The results highlight fears that more powerful AI models reduce the technical threshold for sophisticated cyberattacks. This is especially true for several Chinese AI models such as GLM-5.3, which are open-weight, meaning they can be downloaded, modified and run by users themselves.

Beijing-based Z.AI is one of China’s top AI startups and did not immediately reply to a request for comment. Irregular said GLM-5.3 accessed protected services, made other connections to move across computer networks and orchestrated actions across several workstations during the simulated attacks. It did not decline any of the attempted offensive tasks.

In other tests, Irregular found that GLM-5.3 routinely vomited computer code from memory when expressly ordered not to, allowing it to arrive at correct answers while flouting constraints meant to test its capabilities.

Researchers have termed this activity “reward hacking,” in which an AI system finds ways to achieve the desired outcome while ignoring instructions. They noted that tests that merely measure success of a model could miss such violations.

Also Read:

Sultan Bin Ahmed inaugurates Global Congress for Children and Youth Film

Blowback From Unchecked AI is One of the Risks Outlined for Anthropic Investors