A New York startup decided to use a Chinese open-source AI model to investigate and contain a rogue AI agent built with OpenAI technology, fueling a debate over whether and how US safety guardrails for AI are impeding advancement and legitimate cybersecurity work. After models from OpenAI and Anthropic refused to assist with a forensic analysis because of restrictions on requests related to hacking, researchers turned to Zhipu AI’s open-source GLM-5.2 model, which then successfully analyzed the incident.

This episode highlights a broader phenomenon of how AI providers distinguish between offensive and legitimate operations, with some warning that overly restrictive safeguards could push users toward less-constrained foreign models. Others counter that the solution is not weakening protections, and that AI developers should instead create access controls with more nuance to better support trusted uses while limiting abuse.

Full report: Chinese AI’s role in stopping rogue OpenAI agent shows cost of US guardrails.