Physix Frontier · News Briefing Card (IT Home · Nov 17, 2026)

Alibaba, Tencent, Moonshot tighten cyber capabilities of flagship models

KEY FACTS

  • Alibaba's Qwen4-Max filters attack material from its training data while retaining vulnerability detection capabilities.
  • This version does not release model weights, so users cannot deploy an uncensored version themselves.
  • Tencent removed training data related to exploit generation before the official release of Hy4.
  • Moonshot's Kimi K3.1 Pro removes cyberattack material in two post-training stages.
  • The Ministry of Commerce and other agencies plan to add AI malicious code generation technology to export controls, covering 23 technologies.

KEY DATA

81.0%Kimi K3.1 Pro internal vulnerabilities after filtering
90.5%Kimi K3.1 Pro unfiltered version internal
23Number of AI technology items proposed for export controls
at least 14Number of countries targeted by VAPOR PANDA intrusions

PHYSIX OBSERVATION

Top vendors collectively capping their models' cyber capabilities will in the short term weaken accessibility for some security research and the open-source ecosystem, but this is an inevitable choice under regulatory and public-opinion pressure. The signal worth noting: the open-weight route is contracting, security capability is becoming a hard threshold for model releases, and once the export control list takes effect, AI security will shift from corporate self-discipline to institutional constraint.

Source: IT Home report