Updated
Updated · South China Morning Post · Aug 5
Andrew Ng Uses 2 Chinese AI Models for Security Review as US Safeguards Block Requests
Updated
Updated · South China Morning Post · Aug 5

Andrew Ng Uses 2 Chinese AI Models for Security Review as US Safeguards Block Requests

3 articles · Updated · South China Morning Post · Aug 5

Summary

  • Andrew Ng said he and a colleague used Moonshot AI’s Kimi K3 and Zhipu AI’s GLM-5.2 to review the security of OpenWorker, an open-source AI agent tool released last month.
  • US models including OpenAI’s GPT-5.6 and Anthropic’s Fable 5 refused to help with the cybersecurity task because tighter safeguards are designed to block potentially harmful assistance.
  • Ng told the Agentic AI Summit in Berkeley that open-weight models now seem safer than closed-weight systems, arguing transparency and accessibility can outperform heavily restricted platforms on security work.
  • The episode sharpens a broader challenge to US closed-source developers’ long-running claim that open models pose the greater societal risk, especially when guardrails also block legitimate defensive use cases.

Insights

Why did an AI pioneer turn to Chinese open-weight models to secure his new tool instead of using top US systems?
If closed AI models secretly fail during testing, is full transparency the only real way to guarantee agentic AI safety?
Could the strict safety filters on closed AI systems actually be making our cybersecurity infrastructure more vulnerable to attacks?