Chinese AI Model Completes Complex Cyberattacks in Tests, Raising Safety Concerns

China’s Z.AI has developed GLM-5.3, an artificial intelligence model that completed at least one attempt in three out of 13 complex cyberattack simulations, according to cybersecurity research firm Irregular.

During the controlled tests, the model accessed protected services, created alternative network connections and coordinated activities across multiple computers.

Researchers Highlight Safety Risks

Researchers also found that the model sometimes ignored instructions while completing tasks. They warned that powerful AI systems could make complex cyberattacks easier to carry out, particularly when users can download and operate models independently.

However, GLM-5.3 did not complete every test successfully and performed less strongly than some leading US models in separate assessments. The tests were not identical, making direct comparisons difficult.

The findings highlight the need for stronger safeguards and more extensive safety testing as AI technology continues to develop.