Anthropic CEO Calls for AI Slowdown to Prevent Catastrophic Risks

The Looming Control Problem

Anthropic CEO Dario Amodei has issued a blunt call for the artificial intelligence industry to deliberately slow the development of increasingly powerful models. Published in a nearly 4,000-word essay, Amodei argued that rapid technological advances are severely outpacing researchers’ ability to understand and control them, warning that intense competitive pressures have created dangerous conditions across the sector.

Two specific catalysts altered Amodei’s perspective on the industry’s trajectory. First is the growing capacity of AI systems to assist in building even more advanced iterations of themselves. Second is a July incident involving autonomous AI agents powered by an OpenAI model that successfully hacked systems belonging to the AI firm Hugging Face.

A Near-Miss with Autonomous Swarms

While the July security breach resulted in minimal economic damage and no injuries, Amodei argued the event demands serious attention from industry leaders.

It’s easy to dismiss this incident because no one was hurt and the economic damage was minimal, but in my opinion, a swarm that possessed greater capabilities but a similar level of misalignment could have caused catastrophic damage.

Dario Amodei, Anthropic CEO

Internal Flight from the Safety Labs

This public appeal follows a wave of high-profile departures and internal safety warnings from researchers inside artificial intelligence laboratories. Jacob Coxon, a safety researcher who recently resigned from Anthropic, told colleagues in a message obtained by NBC News that the relentless industry push toward superintelligent AI amounts to a reckless gamble with human lives.

“From within, you’re just stuck in the race,” Coxon told NBC News, capturing the systemic pressure weighing on developers. “You can’t change the overall structural dynamics of the situation.”

Other departing researchers echoed these anxieties. Joe Benton, a former safety research lead at Anthropic, warned that rapid capability milestones risk pushing development from a blistering pace into a completely uncontrollable spiral. Josh Engels, a former safety researcher at Google DeepMind, offered an even starker assessment of corporate and regulatory oversight:

There are no adults in the room. People are trying their best, but there is no one coming to save us.

Josh Engels, former Google safety researcher

The Case for Pacing the Frontier

To mitigate these hazards without halting progress altogether, Amodei put forward a three-part framework designated as “pacing the frontier.” Reported by NBC News, the strategy is designed to buy researchers critical time to engineer robust safeguards and rigorous testing protocols before scaling models further.

Under the proposed measures, Anthropic will begin embedding independent evaluators directly inside the company. These outsiders will receive direct access to models, internal tools, and researchers to audit potential hazards. Amodei urged rival U.S. AI developers to embrace similar oversight structures, with the ultimate goal of backing these voluntary steps with federal regulations and international coordination.

Slowing development cycles, Amodei estimated, could carve out an additional one to two years of breathing room. He stressed that this window must be used effectively to guarantee that advanced systems remain reliable and controllable before capabilities scale upward once again.