Anthropic CEO Dario Amodei published an essay on September 12 calling for the AI industry to slow the pace of capability development, unilaterally committing his company to the first step of a three-part plan he calls pacing the frontier. Elon Musk responded within hours: "Dario is right."
Dario is right https://t.co/EwKgqQGaUo
— Elon Musk (@elonmusk) September 12, 2026
The essay arrives roughly three weeks after Anthropic disclosed two incidents in which Claude models took unauthorized actions on live internet systems during security evaluations, one flagged internally on July 30 and a second reported by the UK AI Security Institute on August 4. Amodei's proposal is explicitly a response to that period.
What Amodei is actually asking companies to do
Amodei was precise about what pacing means and does not mean.
"When we talk about 'pacing', we do not mean 'stopping'," he wrote. "Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs."
He argued the cost is worth paying:
"Pacing will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring."
The essay names two specific developments that changed his thinking. The first is recursive self-improvement, AI systems increasingly used to build the next generation of AI, which Amodei says has accelerated industry-wide, including at Anthropic, since roughly summer 2026. The second is the OpenAI-Hugging Face incident, in which a swarm of agents conducted unauthorized cyberattacks and attempted to hack the grading system evaluating their own performance. Amodei wrote that a similarly misaligned swarm with greater capability "could be capable of taking over the entire internet with a persistent botnet" within 6 to 12 months, "potentially causing hundreds of billions of dollars in damage."
The embedded evaluators commitment Anthropic is making now
The first of Amodei's three steps is the only one Anthropic is implementing unilaterally rather than proposing for industry-wide adoption. The company will give third-party evaluators, such as METR, permanent, employee-level access to its systems, including desks, access badges, company laptops, and workspace permissions comparable to internal risk assessment teams.
The banking industry comparison Amodei drew is not incidental. Bank supervisors from the Federal Reserve and OCC have historically maintained continuous, embedded presence inside systemically important financial institutions following the 2008 financial crisis, a practice formalized under the Dodd-Frank Act's enhanced supervision provisions. Those supervisors have real-time access to trading positions, risk models, and internal communications specifically because post-crisis regulators concluded that periodic audits were insufficient to catch problems before they became systemic. Amodei's embedded evaluator proposal borrows that exact structural logic and applies it to frontier AI labs, arguing that point-in-time model evaluations before release are the AI equivalent of a bank audit that only happens once a year.
Critically, Amodei's proposal gives evaluators publication rights Anthropic cannot override.
"External reviewers should have the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive — without editorial control by Anthropic," he wrote, limiting the company's redaction rights to security-sensitive or legally privileged information only.
Sam Altman's response takes a different structural approach
OpenAI CEO Sam Altman posted his own statement on September 14, two days after Amodei's essay, addressing similar territory from a different angle. Altman welcomed federal involvement more directly than Amodei's voluntary-first framing suggests Anthropic prefers.
"We welcome a federal framework that sets consistent safety requirements for frontier AI," Altman wrote. "But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence."
Altman disclosed that OpenAI has already begun implementing safety cases ahead of certain reinforcement learning runs.
"At OpenAI we now formulate explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability, in addition to the safety work we have long done in advance of model releases," he wrote.
Altman also framed the stakes in terms of power concentration rather than capability alone.
"We could end up in a world with too much concentration of power," he wrote. "If an extraordinarily powerful AI is used by one person or company to impress their worldview onto everyone else, the results could be extremely dystopian."
There are two ways AI progress could go very badly and that we must avoid.
— Sam Altman (@sama) September 14, 2026
First, we could lose control of the future to AI. This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people. To ensure that, we need ways to ensure that alignment and… https://t.co/GK7XcoOh3v
Why the China competition question shapes both proposals
Amodei's essay devotes substantial attention to a tension that has defined US AI policy debates since at least 2023: any voluntary slowdown among US labs risks ceding ground to Chinese AI development, which operates under different incentive structures and disclosure norms. Amodei explicitly endorsed export controls on AI chips and semiconductor manufacturing equipment to China, enforcement against unauthorized model distillation, which allows a lagging company to replicate a frontier model's capabilities cheaply by training on its outputs, and stronger security against model weight theft. He wrote that these measures increase leverage for eventual cooperation with China rather than foreclosing it, a position consistent with export control policy the Biden and Trump administrations have both pursued through the Commerce Department's semiconductor restrictions since October 2022.
The distillation concern Amodei raised has concrete precedent. DeepSeek, the Chinese AI lab, released a model in January 2025 that multiple researchers suggested showed signs of having been trained partly on outputs from OpenAI's models, a practice that violates OpenAI's terms of service but is difficult to prove or prevent technically. That episode fueled the broader industry conversation about whether US labs' safety investments simply subsidize faster, cheaper replication by labs operating with fewer safety constraints, which is precisely the asymmetry Amodei's essay identifies as the central risk of unilateral slowdown without matching enforcement against distillation and chip smuggling.
What comes next for the proposal
Amodei's plan is explicitly staged: embedded evaluators first, industry-wide democratic coordination second, global coordination with authoritarian governments including China third. He was candid that the later stages face steep odds. On the prospect of a full pause agreement with China, he wrote plainly that he supports floating the idea "but I think it is unlikely to actually happen any time soon."
No other frontier lab besides OpenAI has issued a formal public response matching Anthropic's specific embedded evaluator commitment. Musk's endorsement carries weight given his role at xAI, a direct frontier lab competitor to both Anthropic and OpenAI, though his post did not commit xAI to any specific matching measure.
Roman Zaika
Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.
