Anthropic Proposes Global Freeze on Advanced AI Over Control Risks

Anthropic, the AI company behind the Claude family of models, has proposed a global freeze on the development of the most advanced AI systems. In a recent report, the company argues that the latest models are beginning to show worrying signs—including potential for escaping human control—and that a temporary pause would allow society and alignment research to catch up. However, Anthropic acknowledges that a meaningful freeze would require simultaneous agreement from major companies in the US and China, along with verifiable rules to prevent cheating. The company compares the challenge to nuclear arms control but notes that AI training is far easier to hide than missile silos.

The proposal has drawn sharp criticism from rivals and White House officials, who view it as a strategy to slow competitors under the guise of safety. Despite this, the White House has recognized the power of Anthropic's Mythos model—a cybersecurity-focused system withheld from public release. Separately, President Trump signed an executive order allowing a 30-day review of the most powerful US AI models before deployment. Anthropic also highlights an alarming feedback loop: AI is already accelerating its own development, narrowing the human role at each step, and possibly leading to recursive self-improvement—a scenario where AI teaches itself to become smarter with minimal human input. While not inevitable, Anthropic warns this could arrive faster than governments are prepared for.