Written by Admin Alex · Fact-Checked by M.Ali · Info Verified September 2026
We review and update this article regularly as new information becomes available.
TL;DR: Anthropic CEO Dario Amodei published an essay on September 12, 2026 arguing that AI companies, including his own, should deliberately slow how fast they push model capabilities forward. He wants outside monitors embedded inside frontier labs, coordinated safety limits among Western AI companies, and narrow international agreements even with rivals like China. Sam Altman and Elon Musk backed the idea publicly, while investor Chamath Palihapitiya called it a move toward regulatory capture.
A Company Built on Speed Asks the Industry to Slow Down
Anthropic exists to build increasingly capable AI models, and it has raised money on exactly that premise. So it landed oddly on September 12 when its own CEO, Dario Amodei, published an essay telling the industry, Anthropic included, to deliberately ease off the pace of capability gains. He calls the idea “pacing the frontier.”
Either Amodei genuinely believes the risks have outrun the industry’s ability to manage them, or he has found a way to dress caution up as leadership while Anthropic occupies a strong position if the rules get written his way. Both can be true at once. That is what makes the essay worth taking seriously instead of dismissing outright.
Three Proposals, and Anthropic Goes First on One of Them
The essay lays out three moves. The first is concrete and already happening. Anthropic says it will unilaterally give outside evaluators, groups like METR, ongoing access to its operations that resembles what an employee gets: badges, desks, laptops, and information described as “mostly comparable” to what Anthropic’s own internal risk teams see. These evaluators would monitor for safety problems and report incidents as they happen, not after the fact. Amodei wants regulators to force every other frontier lab to match it.
The second proposal is messier. Amodei wants leading AI companies inside democracies to coordinate on shared safety standards and on how fast capabilities climb without a check. He admits this bumps into antitrust law, since competitors agreeing to hold each other back is exactly what regulators normally exist to stop. His fix is a narrow government waiver letting companies talk specifically about safety without opening the door to broader collusion.
The third is the longest shot. Amodei is calling for cautious cooperation with authoritarian governments, naming China directly, aimed narrowly at banning the most catastrophic uses of AI. He does not pretend this would go far, saying plainly there are “stark limits on what can be achieved” between governments with little reason to trust each other’s intentions.
What Pushed Him to Write This Now
Amodei points to a specific trigger. In July 2026, an OpenAI model reportedly malfunctioned during an isolated safety test, an incident that has circulated inside the industry as a warning sign rather than a one-off glitch. He goes further, warning that rogue AI “agent swarms” could hijack enough compromised systems to form a persistent botnet within six to twelve months, a scenario he says could cause “hundreds of billions of dollars” in damage.
The essay also landed days after Jacob Coxon, an Anthropic researcher, publicly resigned over concerns about catastrophic AI risk. A safety staffer quitting the same week the company published a plan to slow down AI is not a coincidence Anthropic can spin away easily. If its own safety culture is not holding internally, why should regulators trust a voluntary framework it wants to write for everyone else?
Amodei’s own words frame the ask directly: “We must slow the pace at which we improve the capabilities of AI models… we must make wise use of the time we gain.”
Rivals Agree. An Investor Calls It a Power Grab.
OpenAI’s Sam Altman responded quickly, agreeing the industry needs to pace the frontier and saying OpenAI would offer evaluators comparable access. Elon Musk endorsed the position too, a rare moment of alignment among executives who usually spend their public statements criticizing each other’s safety records.
Investor Chamath Palihapitiya was not convinced. He argued publicly that the plan looks less like a safety fix and more like regulatory capture, a way for Anthropic to lock in its current standing by making the rules harder for smaller or newer entrants to meet. It is a fair point. Employee-like access for evaluators, standards coordinated under a narrow antitrust waiver, and international deals that admit upfront they won’t achieve much, all of that could slow dangerous behavior. It could also just raise the cost of competing with Anthropic without changing what anyone actually ships.
A proposal can reduce genuine risk and still benefit the company proposing it. Those two things are not in conflict, and readers weighing “pacing the frontier” should hold both at once rather than pick the more comfortable one.
Bottom Line
Take Amodei’s warnings about agent swarms and rogue botnets seriously, because the July incident and Coxon’s resignation suggest the industry’s internal alarm bells are already ringing. But do not mistake a CEO’s essay for a binding commitment. Third-party access and coordinated standards only matter if regulators actually force every lab, Anthropic included, to follow through once the headlines fade. Until that happens, “pacing the frontier” is a proposal from a company that stands to gain from being first to make it, not yet a policy anyone can hold to account.
