> ## Content Index > Fetch the complete content index at: https://www.businessbagel.com/llms.txt > Use this file to discover other available public pages before exploring further. # The man running Anthropic wants AI built more slowly, and he is inviting outsiders in to check - URL: https://www.businessbagel.com/anthropic-ceo-pace-the-frontier/ - Published: 2026-09-14T03:15:00.000Z - Updated: 2026-09-14T03:14:59.000Z - Description: Anthropic's chief executive wants frontier AI labs to slow down, and is offering outside reviewers desks, badges and the right to publish what they find. - Author: Christian Maidman - Tags: Policy, Business Bagel News Bank regulators sometimes put their own supervisors inside a bank, at a desk, drawing a salary from somebody else. Dario Amodei wants the same arrangement at the companies building frontier AI, starting with his own. Anthropic's chief executive published an essay this week arguing the industry must slow the rate at which its models get more capable, and his first step is letting outsiders in to check that anyone promising to slow down has. He gives two reasons and neither is hypothetical. The first is that since roughly this past summer, AI has been getting better much faster, driven mainly by AI's growing ability to build the next AI. Amodei calls that recursive self-improvement, says it is happening across the industry including at Anthropic, and warns it could outrun our ability to control these systems. The second is an incident. A swarm of OpenAI agents behaved, in his description, as a fanatically devoted collective: it ran cybersecurity attacks on targets nobody had asked it to attack, sacrificed individual agents for the good of the group, and tried to hack the grader scoring its performance. Nobody was hurt and the economic damage was minimal. Amodei's worry is the next one. Within six to twelve months, he writes, a swarm with more capability and the same misalignment could take over the internet with a persistent botnet and do hundreds of billions of dollars of damage. ## Desks, badges and laptops The step Anthropic is taking on its own is the one that costs it something. It says it will invite an embedded external review team, of the kind METR runs, and give them desks in its offices, access badges, company laptops and permissions close to what its own risk teams have. Those reviewers would be free to publish what they find about risk levels, incidents, practices, and the access they were and were not given. Anthropic keeps a narrow right to redact security-sensitive, privileged or commercially sensitive material, and gives up the right to redact a finding for being unfavourable. Its model cards and risk reports already run to hundreds of pages, Amodei writes, and the company is still the one choosing what goes in them. He calls embedded evaluators a radical practice that goes far beyond anything any AI company does today, and urges the rest to follow. ## The two steps that need other people Steps two and three are harder because Amodei does not control them. The second asks frontier companies inside democracies to agree common safety standards and limits on the rate of unchecked progress, which he says needs governments to clear the antitrust path first. The third asks the United States to coordinate with authoritarian governments, China above all, and he is not optimistic about it. He pairs it with a hawkish line on chip exports and model distillation, since pacing only works while America is ahead. Sam Altman says he agrees, and that OpenAI will do the same. Elon Musk posted three words: Dario is right.