AINEWS
CEO Dario Amodei turned his call to slow the AI race into a concrete commitment: outside auditors will get employee level access to Anthropic's models, with reporting power over safety incidents.
Anthropic, one of the small group of companies building the most capable AI systems, says it will give outside auditors the same access to its models that its own engineers have. The commitment, announced by CEO Dario Amodei in a Saturday essay titled "We Must Pace the Frontier", converts a rhetorical call to slow the AI industry into a concrete commitment about how a frontier lab proves its safety claims. The new mechanism is permanent third-party access, with reporting power over incidents and over whether the model is being trained to do what its developers intend. The Guardian first reported the plan; Amodei linked the essay on his X account.
The first step of Amodei's three-part plan, which he says Anthropic will "unilaterally" adopt, is to invite evaluators from outside the company in. Those evaluators will, in his words, "verify adherence to our safety measures, report on incidents, and assess models' alignment during training." In plain terms: a third party with the same read-write view of the lab's systems as an Anthropic employee, the right to test the model, and the standing to publish when something goes wrong. The work is continuous, not a one-time audit, and the lab is not choosing the auditor from inside its own stack.
The closest existing example of what that lane looks like in practice is METR, the Model Evaluation and Threat Research group, which on August 26, 2026 published a "brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident". That investigation was retrospective, narrow in scope, and built on cooperation rather than contract. What Amodei is proposing extends the same posture into a permanent, employee-level seat at the table during training itself, applied inside his own company. Whether METR, or any specific evaluator, takes that seat is not yet on the record.
The proposal lands two days after a former Anthropic researcher, Jacob Coxon, publicly quit the lab on Wednesday, warning that AI could precipitate human extinction by 2030. Coxon's critique, which he applied to both Anthropic and OpenAI, was direct: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." He added: "The people building AI earnestly believe that it could kill us all by the end of the decade … No other human activity poses this level of danger." The departure is the proximate trigger for Amodei's essay, and the proposal reads as a partial answer to it. An Anthropic spokesperson told the Guardian the company has "always been transparent that AI will bring both enormous benefits and unprecedented risks" and is building "models with some of the strongest safeguards in the industry."
Amodei hasn't called for a halt. The bolded line in his essay, which he flagged for emphasis, is "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." In the body of the essay he adds that "over the last few months, I have become convinced that fully addressing the risks requires even more prudence — not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up." He is also candid about the pressure: he concedes that a "race to the bottom, spurred by commercial incentives" can make the risks more acute, the same race that a former employee just accused him of helping to run. He does not, in the visible portion of the plan, name which other frontier labs he wants to match the commitment.
Committing to first-step access creates a peer-pressure surface on OpenAI, Google DeepMind, and xAI. Once an Anthropic model is being audited by a third party during training, the public default shifts. Safety becomes a thing a lab demonstrates under external review, not a thing a lab asserts in a blog post. The Verge, the BBC, TechCrunch, and Unite.ai are all running versions of the slowdown story on Saturday; the framing that survives is the one the industry copy machine will repeat.
Two watch items will determine whether Amodei's first step becomes an industry floor or a one-lab gesture. The first is whether the essay's steps two and three, not yet detailed in the publicly available text, address the commercial race Amodei himself names. The second is which outside evaluator, or evaluators, accepts a permanent seat at Anthropic's training runs, and on what publishing terms. Until those two answers land, the proposal is a credible mechanism with a single signatory.