Frontier AI risk is prompting calls for a global pause framework as leaders warn that AI could reach the ability to design its own successors, with coordinated limits needed when governance can’t keep up.

A cross-industry push from Anthropic and OpenAI urges an international, coordinated mechanism to slow or pause frontier AI development when safety and governance lag behind progress.

The broader AI landscape includes rivals like Meta and OpenAI, all racing toward potential superintelligence and spurring safety, monitoring, and governance debates.

A central concern is competitive pressure: if one region slows while others accelerate, slower regions risk losing ground, underscoring the push for universal limits.

Anthropic will convene policymakers, researchers, civil society, and peers to discuss recursive self-improvement, coordination, and safety measures.

OpenAI’s Sam Altman calls for an international AI watchdog, modeled after the IAEA, to set safety standards, monitor advanced AI, and facilitate cooperation.

Anthropic’s Claude is increasingly driving its own development, with researchers noting Claude authored about 80% of the code used by Anthropic as of May, up from earlier levels.

Proposed concrete measures include limits on training run sizes, standardized testing for AI capabilities, and inspection systems that protect sensitive business information while verifying compliance.

Industry reactions include cautious government support for more oversight, with Australia’s AI Safety Institute coordinating with regulators and international partners.

Proposals stress that progress may outpace governance, calling for verifiable, coordinated, time-bound development limits with independent oversight.

Current practices involve safety teams, stress tests, external evaluations, and gradual model releases, but may be insufficient without global coordination.

Anthropic plans a restricted Claude Fable 5 release, signaling tighter controls ahead of broader deployment.