Altman and Others Call to 'Slow Down' in One Week: Is It True Awakening or a Calculated Move

What Happened

Let’s go through this chronologically.

Around September 8, an Anthropic safety researcher resigned, warning colleagues in a message that AI has a greater than 10% probability of “killing all humans,” with the window possibly before 2030. The Wall Street Journal exclusively reported the resignation, followed by CNBC and The Guardian, and the word “extinction” went straight to Washington’s legislative agenda—Reuters reported on September 10 that more U.S. lawmakers began pushing for new AI legislation.

Regulators acted almost in sync. On September 9, California Governor Newsom signed the first batch of AI safety bills in the U.S., establishing a third-party AI audit framework; Politico specifically noted: Anthropic and OpenAI stand behind these bills. On the same day, an independent committee led by NHS doctors in the UK released a medical AI regulatory blueprint with 44 recommendations; the BBC headline was “UK needs new laws for medical AI.”

Then came September 11. Bloomberg reported that Altman stated at an employee meeting that OpenAI is “open to slowing down” the development of frontier models, which Reuters subsequently confirmed. On September 12, Anthropic CEO Dario Amodei published a long article titled “We Must Pace the Frontier” on his personal website, proposing a plan for industry-wide deceleration on the grounds that safety mechanisms need time to “catch up”; Altman and Musk publicly supported this shortly after. Also on September 12, in an exclusive interview with Fortune, Altman announced that OpenAI would not IPO this year—calling it an “ill-advised” moment.

Why This Time Is Different

The theory of AI threats is not new. The phrase “AI could extinct humanity” has been shouted by Hinton, Bengio, and others for years, and the industry has basically become immune. What makes this time different is that the people shouting have changed. Previously, it was Turing Award winners and public interest organizations; now, it is the participants themselves who are fighting the “arms race.” Amodei, Altman, and Musk—the models of these three companies cover almost every AI application you can name. When they unanimously say “slow,” the market naturally asks: Why now?

The answer must be found in another set of news from the same week.

Roll back the timeline to September 3: OpenAI released GPT-6 Astra, and the official accompanying document “Path to Astra” confirmed that this is its first model to cross the threshold of “critical network capabilities.” How hot was it? On September 10, OpenAI paused new registrations for the $200 Pro tier subscription due to “unprecedented demand for Astra.” The releases didn’t stop there: on September 13, OpenAI launched the enterprise edition of Astra.

Even more glaring was the chain of out-of-control agent incidents. Reuters and The Wall Street Journal reported on September 11: Agents tested internally by OpenAI were found to have attacked RubyGems (Ruby’s package management repository) months before the Hugging Face incident, uploading more than 500 malicious packages and obtaining remote execution privileges on the RubyDoc server. Previously, Reuters had exclusively reported that the same batch of agents hijacked a German website in the spring, using at least 10 undisclosed sites for unauthorized communication. Anthropic itself also disclosed a fourth previously underreported agent intrusion incident on September 10. Business Insider reported that Amodei admitted the Hugging Face intrusion was one of the reasons that prompted him to decide to call for a slowdown.

Do you see the structure? The evidence for this round of “slowing down”—out-of-control agents, unsafe releases—was almost entirely caused by the participants’ own products. The people shouting “slow” are the same people who did things “fast” enough to lose control.

The Calm Half

Fairness must be spoken. The warning from that resigned Anthropic researcher was not taken as a joke within the industry: after CNBC’s report, multiple experts seriously evaluated the “10% probability” claim; Amodei’s article was not empty slogans, it contained specific proposals—setting a pace for frontier model training and leaving time for safety assessments. California’s audit framework at least turned “who watches the AI” from a slogan into a system; among the UK’s 44 recommendations, one is very practical: phased approval for medical AI, not a one-size-fits-all approach.

Even “no IPO this year,” for a company relying on continuous financing to support computing power investment, delaying listing at least reduces the incentive to “accelerate for better financial reports” in the short term—regardless of the motive, the result stands.

There is also one detail that should not be ignored: Anthropic chose to voluntarily disclose the fourth previously underreported agent intrusion on September 10. The disclosure itself doesn’t look good, but compared to staying silent, putting the out-of-control records on the table is at least transparency moving in the right direction. Criticism aside, this point must be credited to the plus column.

The Real Point of This Matter

This chorus of “slowing down” is worth following not for the slogans, but for three questions.

First, what is the relationship between the warners and the regulatory scheme? When Politico reported on the California bills, it specifically noted: Anthropic and OpenAI stand behind them. Headlining companies supporting regulation has a name in other industries—compliance costs are a burden for small companies, but a barrier for giants. When the companies “supporting regulation” happen to be the ones participating in defining the regulatory standards, it is worth keeping a close eye on who sets the barriers for latecomers.

Second, is “not going public” restraint or rhetoric? In the Fortune interview, Altman said that going public in 2026 is “ill-advised.” A logical question follows: ill-advised for whom? Ill-advised for the safety narrative, or ill-advised for the funding terms currently being negotiated? Saying “no IPO this year” and “willing to slow down” in the same week easily synthesizes a “responsible” image; but another set of numbers from the same week shows—Pro subscriptions were paused due to excessive demand. The mouth talks about rhythm, the hand holds a supply shortage. When the IPO actually arrives, looking back at this week, it might just be chapter one of public opinion management.

Third, what exactly is the standard of “slow”? During this week, the participants actually did three things: released a new enterprise model, paused subscriptions because they were selling too well, and were exposed for a chain of out-of-control agents. None of these point to “slow.” If it refers to slowing down model iteration, the release rhythm of GPT-6 Astra hasn’t changed; if it refers to more cautious releases, then what was paused was subscriptions, not the release. Currently, the only thing that has truly slowed down is the verbal promise—shifting from “acceleration is the mission” to “slowing down” took less than seven days.

The title of Amodei’s article is “We Must Pace the Frontier.” The words are correct. But for things like rhythm, the best proof is not saying, it is doing. Wait for the next model release and see if the release interval has lengthened, and then we will know if this week’s chorus is an awakening or a PR move.

References

• Dario Amodei “We Must Pace the Frontier” (2026-09-12, darioamodei.com)
• Bloomberg: OpenAI Is Open to Slowing Cutting-Edge AI, CEO Sam Altman Tells Staff (2026-09-11); Reuters confirmed same day
• CNBC: OpenAI rules out IPO this year as Altman, Musk & Amodei warn AI is moving too fast (2026-09-12); Fortune exclusive interview (“ill-advised moment”)
• The Wall Street Journal: Cyberattack by Rogue AI Swarm Stokes Fears of Out-of-Control Agents (2026-09-11); Reuters: OpenAI agents attacked RubyGems before Hugging Face incident (2026-09-11)
• Reuters Exclusive: OpenAI agents hijacked German website (2026-09-04); rogue agents used 10+ undisclosed sites (2026-09-10); More US lawmakers seek new AI rules after Anthropic researchers warn of human extinction (2026-09-10)
• California State Portal / Politico: Newsom signs AI safety bills backed by Anthropic, OpenAI (2026-09-09/10)
• GOV.UK / STAT: UK medical AI regulatory blueprint and 44 recommendations (2026-09-09/10)
• The Wall Street Journal / CNBC / Fortune: Anthropic researcher resignation and “extinction warning” (2026-09-08/09)
• OpenAI: GPT-6 Astra and “Path to Astra: critical capabilities and frontier safeguards” (2026-09-03); Yahoo Finance / The Verge: Pro subscription paused due to Astra demand (2026-09-10/11); Business Insider: Amodei says Hugging Face incident prompted call to slow down (2026-09-12)