Frontier Red Team

Anthropic's frontier red-teaming report claims a popular open-weight Chinese AI model shows what it calls Mythos-class hacking capabilities, alongside weak safeguards. The findings are drawing attention to the risks of freely downloadable frontier models, with debate over whether the security assessment holds up and how open-weight releases should be governed.

Simon Willison highlights Anthropic's new Frontier Red Team, a dedicated team probing advanced AI models for dangerous capabilities and safety failures. The announcement is drawing attention in AI and open source circles, where researchers are weighing what dedicated red-teaming means for how frontier labs assess risks before releasing powerful models.