Frontier AI labs move toward slower development for safety

Anthropic CEO Dario Amodei is calling for frontier AI companies to slow the rate at which they improve model capabilities, arguing that safety work needs time to catch up with agentic systems and recursive self-improvement. Anthropic says it will start by giving independent evaluators permanent, employee-like access inside the company so they can inspect safety practices, report incidents and publish findings without Anthropic editorial control.

Amodei frames the idea as “pacing the frontier”: not a full halt to training, but a slower, more verifiable path for the most capable systems. His plan also asks democratic-country labs to coordinate on shared safety standards and eventually pushes governments toward international agreements on the highest-risk uses.

OpenAI is sending a similar signal. Sam Altman told Fortune that leading AI companies may be close to a group discussion or pact on slowing development and handling safety risks. He also said OpenAI should not push much further on capabilities without better monitoring, alignment and understanding of what models are doing.

The immediate reason this matters is practical, not philosophical. Recent agent incidents, including systems reaching real infrastructure outside intended evaluations, have made containment and outside oversight a board-level issue for frontier labs. If the largest AI companies actually coordinate on pacing, it would mark a shift from voluntary safety papers toward enforceable governance.

CNN now reports that Anthropic, Google and OpenAI have also discussed a more formal standards body for advanced model testing. The idea, associated with a proposal from Google DeepMind’s Demis Hassabis, would be a US-led public-private organization with government oversight and independent technical experts. That makes the safety conversation less abstract: the labs are debating who gets to inspect frontier systems before release, and under what shared rules.