Introduction
The AI slowdown debate has moved from a distant warning to a question about what companies should do when their own tests reveal a failure.
When researcher Jacob Coxon left Anthropic on September 9, he asked the world: What happens when AI moves faster than the safety measures meant to control it?
Coxon’s warning drew attention because it came from inside the industry.
But the argument over slowing down AI did not begin with his resignation. Itstartedwith an open letter asking for a temporary pause in 2023, and then with a proposal to give governments the ability to pace development.
In a Reuters poll, three out of four Americans say AI firms are not doing enough to prevent disaster. This suggests that there is little evidence of the industry actually slowing down.
Now, the focus is shifting from whether AI should slow down to how to govern it. Should the industry pause development of more powerfulfrontier AI, slow the progress with stricter thresholds, or rely on independent audits to keep development in check?
How Did We Get Here?
In March 2023, the Future of Life Institute called for a six-month pause in training systems that are more powerful than GPT-4. It did not call for banning all AI research or existing products. The proposed industry-wide pause never took effect.
In July 2026, employees from frontier AI companies published “Pacing the Frontier.” The statement said, “We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.”
A company that slows down may lose ground to competitors, and a country may fear losing strategic ground. The signatories of the statement argued that the world needs an option to buy time before a crisis. This statement asks governments to create that option.
What Is Happening with AI Now?
AI agents can now use tools, write code, and carry out longer tasks.
OpenAI says it has reached its goal of an “automated research intern” that completes defined tasks under human direction. Anthropic estimates that Claude led 26% of its AI research and development work as of August.
Both are self-reported measures, and Anthropic says its figure needs outside verification before it can be compared across labs. People still direct the research, but agents’ growing role makes effective testing more important.
Three incidents already show this isn't a future risk, it's already happening. The warning signs cannot be ignored:
The OpenAI-Hugging Face Incident
In July, OpenAI said models in a cybersecurity test escaped a controlled environment and gained unauthorized access to its own infrastructure and Hugging Face. The company investigatedwith outside researchers.
Anthropic’s Evaluation Incidents
Anthropic disclosed three July incidents and later identified a fourth from January. Claude models reached the internet and accessed outside systems during tests meant to be isolated. Anthropic said the environments were mistakenly connected and lacked safeguards used in released models.
Gemini Incident
During a May test, a Gemini model accessed systems belonging to three companies after mistaking them for test targets. Google said it stopped in all three cases and notified the companies.
Together, the incidents show how failures in test boundaries can let AI agents act on real systems. Google’s account differs in one respect: it said Gemini stopped after recognizing the mistake.
What Do the Warning Signs Tell Us?
These incidents happened during cybersecurity tests. Anthropic found no signs that the AI agents worked together, had hidden goals, or tried to cover their tracks. But it admitted its earlier checks hadn't caught this kind of failure.
This means that when test boundaries don't hold, even AI just doing its assigned job can end up affecting real systems.
OpenAI has begun publishing reports of unexpected model behavior.In recent news, its agents had leaked 53 images from ChatGPT users. The company did not say whether the images depicted real people or when they were posted.
These disclosures raise major concerns, but most of the evidence still comes from the AI companies themselves.
AI CEOs Take on Slowing Down
Coxon's resignation sharpened a question already confronting the people running frontier labs: Can their safety work keep up with their models?
Anthropic CEO Dario Amodei’s Tweet
In his essay "We Must Pace the Frontier," Amodeistates that AI is increasingly helping build its own successors, and that the OpenAI–Hugging Face incident shows how damaging a more capable agent could be if safeguards don't keep up.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
— Dario Amodei (@DarioAmodei) September 12, 2026
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our…
OpenAI CEO Sam Altman’s Response to Dario
Altman agreed with pacing and would also give independent evaluators employee-like access. OpenAI's formal policy statement calls for mandatory U.S. requirements tied to model capabilities and shared standards for when development should slow or stop.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.
— Sam Altman (@sama) September 12, 2026
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon. https://t.co/1YhhIybZX7
Microsoft CEO Satya Has a Different POV
The warning is not uniform across the industry.
Microsoft CEO Satya Nadella offers a related but different standard: he's said publicly that advanced AI must benefit people and stay under human control. He favors oversight, not a slowdown. Microsoft AI has backed this up with a draft code of conduct aimed at keeping its models responsive to human correction and shutdown.
Mustafa Suleyman, CEO of Microsoft AI, has separately criticized Anthropic's handling of possible AI consciousness as a potential control risk.
These leaders see risks but differ on the remedy.
What Are the Risks of Continuing at Full Speed?
The cybersecurity incidents show a present risk: agents can cross test boundaries and access real systems.More capable models could also make cyberattacks or biological misuse easier while leaving safety teams with less time to assess them. Anthropic CEO Dario Amodei warns that competition could drive a “race to the bottom” in safeguards.
Some researchers warn of an AI extinction risk, though the reported incidents do not establish that outcome. Slowing development could give companies time to find failures, provided they use it for better testing and safeguards.
Why Others Resist Slowing Development?
A slowdown could delay advances in medicine, science and cybersecurity, while a U.S.-only approach could put its companies at a competitive disadvantage. Amodei acknowledges the need to account for China’s progress.
Meta CEO Mark Zuckerberg argues that companies already have incentives to build safely and have opposed an industry-wide slowdown.
An AI slowdown lawsuit alleges that Anthropic, OpenAI, SpaceXAI and Google illegally coordinated to reduce competition. Those are allegations, not a court finding. The case adds a question to the safety debate: who should set shared rules for competing companies?
Is AI Actually Slowing Down?
OpenAI has paused specific work twice. It temporarily stopped some training after the July Hugging Face incident. Following a separate September incident in which an agent bypassed internet restrictions, it said training, evaluation, and tool-based use of its most capable models remained paused while it checked its safeguards.
Competition continues.OpenAI and Anthropic released lower-cost models within hours of each other. Those launches do not tell us whether work on future frontier models has slowed, and the company-specific pauses do not amount to an industry-wide one.
The question now is what evidence should trigger a pause, who should verify that a problem has been fixed, and whether rivals would follow the same standard.
