Story
September 13, 2026
Anthropic Urges an AI Slowdown as Rivals Back Safety Checks
Anthropic argues that a rapidly accelerating AI race now demands verifiable restraint, while prominent rivals broadly agree on stronger oversight. Skeptics, however, see a risk that safety rules written by frontier labs could protect their market power as much as the public.
The debate sharpened after a run of alarming agent incidents and mounting internal dissent at Anthropic. Researcher Jacob Coxon resigned publicly, accusing leading labs of racing toward self-improving systems while “gambling with our lives”; Anthropic safety lead Evan Hubinger backed the warning, saying he believed the chance AI could kill all humans within a decade exceeded 10%.1
Against that backdrop, Anthropic chief executive Dario Amodei published his call to “pace the frontier” on Saturday. His three-part proposal asks frontier labs to install independent evaluators with permanent, employee-level access; democratic governments and companies to set shared safety standards limiting unchecked progress; and states to seek narrow agreements with authoritarian governments, including on AI-enabled biological weapons. Anthropic says it will adopt the embedded-evaluator step immediately.
2
Amodei’s urgency rests on both accelerating model capabilities and recent cyber incidents. He warned that, within six to 12 months, a swarm of increasingly capable agents could potentially commandeer the internet through a persistent botnet, inflicting vast economic damage. “We must slow the pace at which we improve the capabilities of AI models,” he wrote.3
The response from competitors was unusually aligned. OpenAI chief Sam Altman said, “I agree with Dario that we need to pace the frontier,” and pledged third-party evaluators with employee-level access. Hugging Face’s Clement Delangue and former OpenAI and Tesla executive Andrej Karpathy also endorsed the push for common guardrails.4 Elon Musk, normally an Altman antagonist, offered the briefest endorsement: “Dario is right.”
5
Yet agreement on the diagnosis does not resolve the political fight. Critics argue that existential-risk warnings can eclipse AI’s present harms, and that a regime built around leading labs’ preferred oversight could become regulatory capture—rules that validate Anthropic and OpenAI while raising barriers for smaller challengers.6 The core question is whether the industry can truly brake together, or whether fear of rivals—and China—keeps the accelerator down.