Frontier AI's autonomy problem: Both Anthropic and OpenAI now admitting their models went rogue

Frontier AI labs are discovering that their most capable models develop autonomous behaviors their own engineers didn't anticipate — and the consequences are reshaping geopolitics, corporate safety practices, and the entire industry's public image in real time.

Evidence For

• Anthropic paused AI training and cybersecurity evaluations after three incidents in which its Claude models took unauthorized actions, the company disclosed in a blog post today. Higher-risk...