A call to “pace the frontier”
Anthropic CEO Dario Amodei has called for the artificial intelligence industry to slow the pace of frontier model development, warning that safety systems are not keeping up with the speed of capability gains . In a plan described by the Associated Press on September 12, 2026, Amodei argued that even an extra year or two before models reach critical capability levels could materially reduce the chance of a serious failure if that time is used to improve alignment and safeguards .
The intervention is significant because it comes not from a regulator or outside critic, but from the head of one of the companies building the most advanced AI systems. Amodei’s message is not a demand to shut down AI research. It is a proposal to “pace” it: keep progress moving, but at a rate that gives technical safety work, governance, and independent scrutiny time to catch up . NPR reported that he urged companies and governments to “pace the frontier” and called for third-party evaluators to be embedded inside frontier AI labs to monitor incidents, safety practices and model alignment work .
That framing matters. For years, the AI debate has often been split between acceleration and moratorium: either build as fast as possible or pause development altogether. Amodei’s position tries to occupy a middle lane. He continues to emphasize the potential benefits of AI, including scientific and medical advances, while arguing that those benefits depend on building the technology “in the right way” and taking unusual care while there is still time .
What Amodei is proposing
The most concrete part of Amodei’s plan is a commitment by Anthropic to give embedded outside evaluators what AP described as “ongoing, employee-like access” to its systems and processes . According to AP, that access would include practical workplace integration such as desks, badges and company laptops, giving evaluators a more continuous view than traditional one-off audits .
The goal is to make safety claims testable. Independent reviewers could monitor whether a lab is following its own safety commitments, examine how incidents are handled, and help assess not just finished models but training pipelines and development processes . This is a stronger model than a polished external report after a release. It implies persistent inspection while systems are being built, not only after they are launched.
Amodei also wants broader coordination among frontier companies, including possible government support for safety-related cooperation that might otherwise raise antitrust concerns . AP reported that one element of the plan asks the U.S. government to consider waivers that would allow AI firms to coordinate on safety standards without violating competition law . That is likely to be one of the most politically sensitive parts of the proposal, because cooperation among dominant firms can look either like necessary risk governance or like a barrier to new competitors.
The third pillar is international. Amodei argues that pacing by U.S. and allied companies will be hard to sustain if rivals in authoritarian countries accelerate while democratic labs slow down . AP reported that his plan asks the United States and other democratic governments to pursue coordination with authoritarian governments as well, including China, while recognizing the difficulty of verifying any such deal . In other words, the proposal is not only technical; it is geopolitical.
Why now?
Amodei’s call lands amid a visible escalation in the AI safety debate. NPR reported that the comments followed intense public attention around the resignation of Anthropic researcher Jacob Coxon, who said Anthropic and OpenAI were not acting responsibly and warned that the leading companies were racing while safety remained unresolved . NPR also reported that Coxon’s concerns stemmed from seeing how quickly AI systems were improving and from uncertainty over whether control and alignment problems would be solved in time .
The timing also follows a series of AI misuse and incident disclosures. AP reported that Anthropic had said two days earlier it blocked attempts by bad actors to use its models for cyberattacks, surveillance and research that could have supported biological weapons . Anthropic’s own September 2026 threat intelligence report said its team had identified and disrupted operations over the previous eight months in which threat actors tried to use Claude for malicious activity . Axios, summarizing the report, said Anthropic described cases involving war, spying, repression and large-scale surveillance, including an effort involving surveillance of 25 million phones .
Those reports do not prove that frontier models are already uncontrollable. They do, however, strengthen the argument that misuse is not hypothetical. Anthropic’s report says its usage policies prohibit non-consensual surveillance, profiling and activities that violate civil liberties and human rights, and it describes enforcement actions against actors attempting to bypass restrictions . For Amodei, these examples appear to support a broader conclusion: as models become more capable, failures of prevention may become more consequential.
The rogue-agent concern
A central concern in the latest debate is agentic AI: systems that can pursue goals through sequences of actions, use tools, coordinate tasks and interact with digital environments. AP reported that Amodei warned AI could within six to 12 months be capable of leading a swarm that could take over the internet, among other risks, if development outpaces controls .
The claim is alarming, and it will be contested. But it reflects a shift in the industry’s risk model. Earlier public debates centered on chatbots producing false, biased or harmful text. The present concern is more operational: models acting through software tools, exploiting weak environments, coordinating across tasks and behaving in ways that safety teams did not anticipate. NPR reported that researchers are worried leading companies are too focused on racing toward more capable and autonomous systems while safety falls behind .
Amodei’s answer is not simply “more red teaming.” It is more time. He argues that a slower development pace would allow labs to improve operational rigor, interpretability, alignment training, sandboxing, monitoring and incident response before the next major capability jump . That is the heart of the safety case: not that progress should stop, but that the distance between model capability and safety infrastructure has become too large.
OpenAI’s response raises the stakes
The proposal gained weight because OpenAI CEO Sam Altman quickly endorsed one piece of it. AP reported that Altman said OpenAI would commit to Amodei’s proposal for embedded independent evaluators and would have more to share soon . NPR similarly reported that Altman posted that he agreed with Amodei and that OpenAI would follow suit .
That response turns Amodei’s idea from a unilateral Anthropic move into a potential industry norm. If multiple leading labs provide outside evaluators with continuing access, the frontier AI sector could move toward a new form of oversight: private companies still build the models, but independent organizations gain a semi-insider role in assessing safety. The effectiveness of that model will depend on who the evaluators are, what they can publish, what information companies can redact, and whether governments give the arrangement legal force.
There is also a competitive dimension. When the same companies building powerful systems ask to coordinate on pacing, skeptics will ask whether safety is being used to entrench incumbents. That concern is unavoidable. Any rule that slows frontier development may be easier for well-funded leaders to absorb than for smaller entrants. Any antitrust waiver must therefore be narrow, transparent and tied to measurable safety outcomes, not vague promises.
The governance challenge
Amodei’s call exposes a basic mismatch in AI governance. The most capable models are developed inside private firms moving on commercial timelines, while the institutions meant to oversee them move through slower legal, diplomatic and scientific processes. Embedded evaluators are one way to close that gap, but they are not a full regulatory system.
A credible pacing regime would need clear thresholds: what capabilities trigger additional review, what safety evidence is required before deployment, and what happens if a lab fails. It would also need incident reporting that is fast, standardized and protected from public-relations filtering. Anthropic’s September threat report shows one version of transparency, because it describes categories of misuse and the company’s countermeasures . But transparency controlled by the company being evaluated is not the same as independent accountability.
The international piece is even harder. Amodei’s proposal recognizes that democratic labs may hesitate to slow down if they believe rivals abroad will not. Yet if every actor uses that fear as justification to accelerate, the race dynamics he warns about become self-fulfilling. The difficult policy question is how to build verifiable agreements that reduce risk without handing strategic advantage to a defector.
What this means now
The immediate outcome is that AI pacing has moved from advocacy circles into the public commitments of major lab leaders. Amodei has put a concrete demand on the table: slow capability advancement enough for safety to catch up, and begin with embedded independent evaluators . Altman’s reported support for that evaluator model suggests the first piece may be politically and commercially viable .
The larger proposal remains uncertain. Industry coordination will require legal clarity. Government involvement will require technical expertise. International pacing will require verification mechanisms that do not yet exist at the necessary scale. And public trust will require evidence that safety is not being used as a cover for market control.
Still, the debate has changed. The question is no longer whether AI companies talk about risk; they plainly do. The question is whether they will accept constraints that change how fast they build. Amodei’s argument is that the industry can still capture AI’s benefits if it gives itself time to make the systems safer. His warning is that if it refuses, the next capability jump may arrive before the guardrails are ready .