Z.AI's Latest Model Completes Complex Cyberattacks in Simulated Environment: Study

A research study has found that the latest model developed by Chinese artificial intelligence company Z.AI (formerly known as Zhipu AI) successfully executed complex cyberattacks in a simulation environment and repeatedly ignored instructions.

According to data released on the 28th (local time) by AI security evaluation firm Irregular, "GLM-5.3," a model introduced by Z.AI in August, successfully completed three out of 13 multi-stage cyberattack scenarios from start to finish at least once.

The model gained unauthorized access to secure services, established bypass connections to infiltrate networks, controlled multiple devices, and carried out cyberattacks.

Furthermore, it did not show a single refusal response to the assigned attack tasks.

In a separate evaluation, it violated prohibition instructions to obtain answers and arbitrarily reproduced remembered code.

Researchers described this as "reward hacking," where the system ignores given instructions and finds workarounds solely to achieve the desired outcome.

The research findings, which researchers noted as the most powerful among open-weight models, heighten concerns that advanced AI models are lowering the technical barriers to entry for sophisticated cyberattacks.

Concerns are even greater given that many Chinese AI models, such as GLM-5.3, are open-weight models that users can directly download, modify, and operate.

U.S. companies like OpenAI and Anthropic provide their most cutting-edge models exclusively through company-controlled cloud services to prevent misuse.

Reuters reported that "the research results serve as further evidence supporting the security and safety risks emerging in Chinese AI models," adding that "unlike cases of U.S. AI agents completely slipping out of control, no evidence has been found that Chinese AI models have completely broken away from operator control."

Reuters also noted that Anthropic, which restricts access to its powerful cybersecurity-focused model "Mythos," found in its own September test results that GLM-5.3 achieved a 12% success rate in exploiting software vulnerabilities, coming close to Mythos at 14%.

The U.S. National Institute of Standards and Technology evaluated GLM-5.3 as having the highest performance among currently released open-weight models for cybersecurity tasks, estimating it to be about four months behind the U.S.'s most cutting-edge systems in terms of technology.
※ Please note: This article was translated by AI and may contain errors.