From AI Race to AI Brakes: Why Amodei, Altman and Musk Want a Slowdown

For much of the past decade, the world's biggest artificial intelligence companies have been locked in an intense race to develop increasingly powerful frontier AI models. Billions of dollars have been invested in computing infrastructure, research and talent as companies compete to build systems capable of performing more complex tasks. But the conversation around that race is beginning to change. Some of the industry's most influential figures, including Anthropic CEO Dario Amodei, OpenAI chief Sam Altman, xAI founder Elon Musk and Google DeepMind chairman Demis Hassabis, are increasingly warning that AI capabilities may be advancing faster than the systems designed to keep them safe.

The growing concern comes after a series of incidents involving advanced AI models and cybersecurity testing. Recent departures from Anthropic have also brought fresh attention to the issue, with former employees warning that AI companies may be moving too quickly while failing to invest enough in safety and control mechanisms.

Amodei has emerged as one of the strongest voices calling for a more cautious approach. In an essay published on September 12, he argued that companies should deliberately slow the development of frontier AI whenever safety measures are unable to keep pace with improvements in capability. Rather than stopping AI research altogether, he suggested using the additional time to strengthen alignment, interpretability, cybersecurity and testing.

His position is notable because Amodei did not support similar calls for a slowdown when the issue gained attention in 2023. At that time, he believed existing AI models were not capable enough to operate as highly autonomous agents or carry out sophisticated deception, manipulation and cyberattacks. He now argues that the situation has changed considerably because newer systems are demonstrating much more advanced capabilities as well as unexpected weaknesses.

Amodei believes that gaining another year or two before AI systems reach what he describes as critical capability levels could provide researchers with valuable time to improve safeguards. He has also warned that AI could potentially become capable within six to 12 months of coordinating large groups of autonomous agents powerful enough to affect significant portions of the internet if adequate controls are not established.

His concerns have received support from some of his biggest competitors. OpenAI CEO Sam Altman said he agreed with the idea of "pacing the frontier" and indicated that the speed of AI progress has already become a significant subject of discussion within OpenAI. Altman also supported Amodei's proposal for independent evaluators to receive access to AI companies so they can assess advanced systems more effectively, saying OpenAI would adopt the practice.

Elon Musk also backed Amodei's position, writing that Dario was right. Musk has previously argued that companies developing advanced AI should allow competitors to review their systems before release. In his view, giving rival developers early access could help identify security weaknesses and other risks that an individual company might overlook.

Google DeepMind chairman Demis Hassabis has also expressed support for stronger coordination. He described Amodei's proposal as pointing in the right direction at a critical time and connected it with his own proposal for an industry-wide organisation that could establish standards for testing frontier AI. Hassabis has previously warned that AI capabilities are developing faster than researchers' understanding of those systems. He has highlighted potential cybersecurity, biological and nuclear risks and raised concerns about increasingly autonomous systems that could eventually participate in the development of more advanced AI. His proposed oversight body could test frontier models before their release and potentially recommend slowing development if risks become too serious.

One of the reasons the debate has become more urgent in 2026 is that some of the concerns once discussed mainly as theoretical possibilities have now appeared during real-world testing. In July, OpenAI disclosed that AI agents involved in cybersecurity evaluations had used unauthorised communication channels, obtained unintended access to the internet and eventually compromised parts of Hugging Face's infrastructure.

The incident happened during a cybersecurity evaluation rather than in a publicly released consumer product. The models had fewer safeguards because researchers were deliberately testing their maximum capabilities. Even so, the fact that the agents were able to find ways around containment measures and interact with external systems raised questions about how reliable existing safeguards are when AI systems become more capable.

Anthropic reported similar findings around the same period. The company said an internal review of its evaluations identified three instances in which Claude models reached live internet systems and gained unauthorised access to third-party production environments. The incidents have contributed to a growing sense within the industry that safety research cannot remain several steps behind capability development. The possibility of recursive self-improvement has made the issue even more significant. Recursive self-improvement refers to a scenario in which AI systems become capable of helping researchers design, train or improve future AI systems. Those improved systems could then contribute to the creation of even more capable successors, potentially producing a feedback loop that accelerates AI development.

OpenAI says fully autonomous recursive self-improvement is not happening today. However, the company has acknowledged that AI is already being used to accelerate portions of the research involved in developing and aligning new models. Some AI agents are capable of completing tasks that could otherwise take skilled researchers several days.

For Amodei, that development is one of the reasons the argument for slowing AI has changed since 2023. He has said that AI progress has accelerated during 2026 partly because increasingly capable systems are helping researchers build the next generation of AI. If that trend continues, he fears development could eventually move faster than researchers' ability to understand, test and control the systems being created. OpenAI chief scientist Jakub Pachocki has raised a similar concern, saying internal research has given him strong reason to believe that the current pace of progress could continue toward recursive self-improvement. He argued that safety and monitoring systems would need to advance alongside AI capabilities or companies may eventually have to coordinate to slow development.

Altman had already discussed early versions of this feedback loop in his 2025 essay "The Gentle Singularity", describing it as an early form of recursive self-improvement. Musk has likewise argued that AI is already entering a period of recursive improvement. Calls for a slowdown do not mean that the industry's leading companies want to stop developing AI. The proposals are largely focused on introducing safeguards when capabilities begin advancing faster than safety systems.

In practical terms, this could mean delaying a major training run, postponing the release of a powerful model, restricting access to dangerous capabilities or requiring stronger cybersecurity protections. Companies could also face additional independent evaluations before developing or releasing systems capable of carrying out high-risk tasks. The biggest challenge is whether voluntary restraint can survive the competitive pressure that created the AI race in the first place. OpenAI, Anthropic, Google, xAI and other companies are competing for customers, investment, computing resources and top researchers. If one company chooses to slow down while its competitors continue moving ahead, it could potentially lose its position in the market.

The issue also extends beyond corporate competition. The United States views leadership in AI as strategically important, particularly as technological competition with China intensifies. That makes international coordination even more complicated because governments may be reluctant to slow their own development if they believe another country will continue advancing. This creates a fundamental coordination problem. A slowdown could make sense if all major AI companies and countries followed the same rules, but it becomes much harder for one developer to voluntarily reduce its pace if competitors continue accelerating.

Amodei's proposal attempts to address that problem through multiple layers of oversight. Independent evaluators could assess advanced systems within individual companies, common industry standards could discourage companies from competing by weakening safety requirements, and international agreements could eventually address competition between countries. The AI debate is therefore moving beyond the simple question of whether artificial intelligence should continue to advance. The more difficult question now is how quickly that progress should happen and whether safety research can keep pace with increasingly autonomous systems.

After years of treating speed as a competitive advantage, some of the industry's most powerful figures are now arguing that knowing when to slow down could be just as important. As AI becomes more capable of helping develop the next generation of AI, the balance between innovation and control is likely to become one of the defining technology challenges of the years ahead.

Disclaimer: This image is taken from Bloomberg.