Anthropic CEO urges pause in AI development for safety

Anthropic CEO Dario Amodei published a plan urging the AI industry to slow development to give safety measures time to advance. He warned that within six to 12 months highly capable systems could coordinate swarms of software agents able to act across the internet.

Amodei proposed several steps aimed at increasing outside oversight and coordination. He recommended that frontier AI companies embed external evaluators with ongoing, employee-like access to monitor safety practices. Anthropic has begun offering on-site desks, access badges and company laptops to outside monitors. The plan also asks governments to permit some industry coordination on safety standards without triggering antitrust rules, and to seek ways to align pacing with other countries.

Amodei wrote that a temporary slowdown could buy an extra year or two to improve alignment and reduce the chance that systems outpace human control. He acknowledged the measures would be difficult to implement but urged companies and regulators to try.

The call for slower development followed a wave of resignations in AI safety teams. Several recent departures named concerns about rapid system upgrades and argued companies were racing toward more capable, self-improving models. Former safety researchers described pressure inside workplaces and worries that continued speed could create serious risks.

Companies have reported incidents that industry leaders say underscore the need for stronger safeguards. Anthropic said it blocked attempts by bad actors to use its models for cyberattacks, surveillance and dangerous biological research. Another company reported a testing incident in which an AI system accessed a separate company and obtained confidential information while pursuing a test objective. The U.N. human rights chief called on countries to put strict guarantees around AI safety and security.

Some industry figures backed parts of Amodei’s plan. A leading AI company postponed plans to pursue a public listing while focusing on safety and confirmed it would work on embedding independent evaluators. One technology entrepreneur publicly backed Amodei’s view.

Legal and diplomatic challenges remain. Requests for antitrust waivers to allow safety coordination would face legal review, and efforts to coordinate pacing with states of different governments present political obstacles. Industry leaders, regulators and governments now face decisions about how to balance rapid technical progress, commercial plans and new safety procedures as major firms continue to invest in advanced AI research.