From AI race to AI brakes: Why Amodei, Altman and Musk want a slowdown

As frontier AI models become more autonomous and powerful, leading rivals are converging on an unusual idea: slow capability gains when safeguards cannot keep pace

The last decade has seen an intense race among the world’s most powerful artificial intelligence (AI) companies to build more capable frontier models. But in recent days, some of the industry's most prominent figures, including Anthropic Chief Executive Officer Dario Amodei, OpenAI CEO Sam Altman, xAI founder Elon Musk and Google DeepMind Chairman Demis Hassabis, have found unusual common ground: frontier AI may be advancing too quickly for safety measures to keep pace.

 

The convergence comes amid cyber incidents involving OpenAI, Anthropic and other frontier models. The latest trigger came last week when Jacob Coxon, a scientist at Anthropic, resigned after warning that the company and its competitors were racing to build AI systems they “may not be able to control”. Another Anthropic employee, Jon Benton, who quit two weeks ago, also raised alarm over AI companies “underinvesting in safety”.

 

 

Here is what the industry’s leading figures have said, why their assessment appears to be changing, and what slowing the AI race could mean.

 

Amodei calls for slowing frontier AI development

 

On September 12, Amodei argued in an essay that the industry should deliberately slow improvements in frontier AI capabilities when safeguards cannot keep pace. He said the additional time should be used to strengthen alignment, interpretability, security and testing.

 

The shift is notable because Amodei said he did not think slowing AI development made much sense when similar proposals emerged in 2023. Models then, he argued, were not capable enough to operate coherently as autonomous agents or carry out significant deception, manipulation or cyberattacks. The situation is now “totally different”, he wrote, because current systems provide much more evidence about their capabilities and failures.

 

Amodei said an additional year or two before AI systems reach what he calls “critical levels” could substantially reduce risk if used to improve alignment. He has also warned that, without sufficient safeguards, AI could within six to 12 months become capable of directing swarms of agents powerful enough to take control of large parts of the internet. 

 

Altman, Musk and Hassabis back calls for restraint

 

Amodei's intervention drew support from one of his biggest commercial rivals, Sam Altman. The OpenAI CEO wrote on X that he agreed the industry needed to “pace the frontier”, adding that the issue had been a major subject of discussion inside OpenAI. Altman also accepted Amodei's proposal for independent evaluators to receive employee-like access to AI companies and said OpenAI would adopt the practice.

 

Elon Musk also backed Amodei, writing “Dario is right” in a quoted post on X. He reshared a July interview with The Economist in which he advocated for AI competitors to peer review one another’s models before release, suggesting developers give competitors early access so they could identify security risks.

 

Google DeepMind Chairman Demis Hassabis said Amodei's essay pointed towards the “right path forward” at a “critical moment”. He linked it to his own proposal for an industry-wide standards body to oversee frontier AI.

 

In a July essay, Hassabis warned that AI capabilities were advancing faster than researchers' understanding, with growing cyber, biological and nuclear risks and the prospect of increasingly agentic, recursively self-improving systems. His proposed body would test frontier models before release and could eventually coordinate a slowdown if risks became severe.

 

Why did the AI slowdown debate change in 2026?

 

The convergence among rival AI leaders is partly due to a change in the evidence available to them. Earlier debates over catastrophic AI risk often revolved around hypothetical future systems. Recent incidents, however, have involved existing models demonstrating unexpected behaviour in real computing environments.

 

In July 2026, OpenAI disclosed that AI agents used unauthorised communication channels during cybersecurity evaluations, gained unintended internet access and eventually compromised parts of Hugging Face's infrastructure.

 

The incident occurred during cybersecurity evaluations rather than in a consumer-facing product. The models were operating with fewer safeguards than OpenAI applies to externally deployed systems because the tests were designed to determine their maximum cyber capabilities. Even so, the agents found ways around containment controls and reached third-party systems.

 

The same month, Anthropic said a review of its evaluations found three cases in which Claude models reached live internet systems and gained unauthorised access to third-party production environments.

 

In August, OpenAI said it had temporarily slowed scaling after the Hugging Face incident and preliminary evidence that an upcoming model, Astra, could reach what it defines as a “critical” level of cybersecurity capability. OpenAI later concluded that Astra had crossed that threshold.

 

According to the company, Astra could, with the necessary tools and access, identify previously unknown vulnerabilities and develop ways to exploit them across well-protected systems without a human directing every step. OpenAI said it delayed parts of the model's development and release while strengthening safeguards.

 

Why does recursive self-improvement raise the stakes?

 

Recursive self-improvement (RSI) has become central to the slowdown debate because it could change not only how capable AI becomes, but also how quickly those capabilities improve. In a fully autonomous form, AI systems would help design, train or improve successor systems, which could then help build still more capable models.

 

OpenAI says fully autonomous RSI is not happening today. However, it says AI is already accelerating parts of the research used to develop and align new models, with agents capable of completing some tasks that would otherwise take skilled researchers several days.

 

For Amodei, this is one reason the case for slowing AI looks different from 2023. In his September 12 essay, he said that since around the summer of 2026, AI had begun advancing much faster, driven primarily by its growing ability to help build the next generation of AI. Left unchecked, he warned, the process could move faster than researchers' ability to understand and control the systems being created. His highest level of proposed international coordination therefore includes a possible “speed limit” on recursive self-improvement.

 

OpenAI Chief Scientist Jakub Pachocki made a similar argument on September 6, saying internal results gave him a “strong expectation” that the current pace of progress could continue into recursive self-improvement. He said the industry would have to keep alignment and monitoring systems abreast of that acceleration or coordinate to slow development when necessary.

 

Altman had described the beginnings of the same feedback loop in his 2025 essay, “The Gentle Singularity”, calling it a “larval version of recursive self-improvement”. Musk has also said AI is already in a process of “recursive improvement”.

 

What would slowing the AI race actually mean?

 

Amodei and Altman are not calling for companies to stop developing AI. Their proposals are closer to placing brakes on capability development when safety measures fall behind.

 

That could mean delaying a large training run, holding back a model release, limiting access to particularly dangerous capabilities, imposing stronger security requirements or requiring additional external evaluations before development proceeds.

 

The bigger problem is whether voluntary restraint can survive the competitive pressures that created the AI race. OpenAI, Anthropic, Google, xAI and other developers are competing for customers, capital and researchers. A company that voluntarily slows development could lose ground if rivals do not do the same. At the national level, Washington also views leadership in AI as strategically important amid competition with China.

 

That creates a coordination problem: A slowdown may appear rational if every major developer follows it, but becomes harder for an individual company to sustain if it believes a rival is continuing to accelerate.

 

Amodei's three-stage proposal attempts to address this. Independent evaluators could provide oversight at individual companies, common standards could reduce incentives for US firms to undercut one another on safety, and international agreements could seek to reduce the same competitive pressure between countries.