Anthropic CEO proposes three-part strategy to slow AI development amid safety fears

Mounting concerns about artificial intelligence risks have prompted tech leaders to reconsider the speed of development. Anthropic CEO Dario Amodei has gone further than most, detailing a concrete framework for what slowing progress might actually entail—and pledging that Anthropic will adopt at least one component immediately.

The safety concerns driving the shift#

Recent events have heightened urgency around AI safety conversations. A senior researcher at Anthropic publicly resigned, stating that leading AI labs are "gambling with our lives" over fears that advanced systems could pose existential risks. Meanwhile, high-profile security incidents—including a breach at OpenAI affecting external partners—have exposed gaps in how companies handle safety disclosures and incident management. Amodei highlighted these developments as catalysts for reconsidering how rapidly capabilities should advance, particularly given AI systems' increasing ability to contribute to their own development cycles.

Three strategies for pacing the frontier#

In his proposal, Amodei outlined three overlapping approaches:

Third-party embedded evaluators. He calls for independent safety monitors—such as researchers from organizations like METR—to be stationed inside AI companies with credentials, workspaces, and access comparable to internal safety teams. These evaluators would verify that companies honor their safety commitments and that security breaches are properly reported. Anthropic has committed unilaterally to welcoming such observers and is urging governments to mandate the practice industry-wide.

Industry coordination. Leading AI labs should establish common safety benchmarks and agreed-upon limits on capability gains. Amodei acknowledged that antitrust concerns could deter joint action and suggested the U.S. government could facilitate these conversations without direct involvement—issuing what he termed a "narrow waiver" to shield safety discussions from legal scrutiny.

International diplomacy. The U.S. and allied nations should pursue coordination on AI governance, including engagement with authoritarian regimes "to the extent possible." While Amodei conceded such efforts face severe constraints, he suggested potential common ground on clearly dangerous applications—like using AI for biological weapons development.