Anthropic Proposes Verifiable Global Pause on Cutting‑Edge AI

Overview

• Anthropic published a detailed note Thursday that used internal metrics to warn that trends could make recursive self‑improvement plausible and reported that Claude now writes a large share of the company’s integrated code.
• The paper calls for a multilateral, verifiable mechanism allowing leading developers to slow or temporarily suspend work and says Anthropic would slow its own development if other top labs adhered in a provable way.
• Designing credible verification tools is technically difficult because AI work can be hidden and the proposal must overcome strong political resistance driven by national security and fears of giving China an advantage.
• Critics and some U.S. officials say the warning may serve public relations or IPO aims, noting Anthropic has limited distribution of its Mythos model and is taking steps toward a public offering.
• Anthropic plans to convene regulators, researchers and industry to build verification methods and governance options, but the effort faces steep practical and geopolitical hurdles that will shape regulation and safety research.