Increasingly capable AI systems have intensified concerns over cybersecurity, alignment and the ability to keep advanced models under human control. #
The leaders of several of the world's most powerful artificial intelligence companies are publicly backing a slower pace for frontier AI development, marking an unusual point of agreement among executives whose companies have spent years competing to build increasingly capable models.
Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, Elon Musk and Google DeepMind CEO Demis Hassabis all endorsed some form of pacing for advanced AI development on Saturday. The endorsements came within hours of one another after Amodei published a detailed proposal arguing that AI capabilities are advancing faster than the industry's ability to understand and control the risks.
In his essay, "We Must Pace the Frontier," Amodei said companies should slow the rate at which they improve the capabilities of their most advanced models, while continuing AI research and development. He argued that additional time was needed for alignment research, security measures, monitoring and outside evaluation to keep pace with increasingly powerful systems.
Amodei proposed a three-part approach beginning with embedded third-party evaluators inside frontier AI companies. Anthropic said it would unilaterally commit to that step, which would give outside evaluators ongoing access to assess safety practices, training processes and potential alignment problems.
The second stage would involve AI companies in democratic countries coordinating on common safety standards and limits on unchecked development, with government support where necessary. A third stage would seek international coordination, including with China, although Amodei acknowledged that reaching and verifying such agreements would be substantially more difficult.
Musk backed Amodei shortly after the essay was published, writing on X that "Dario is right." Altman followed by saying he agreed that the industry needed to "pace the frontier," while Hassabis later said Amodei's proposal pointed in the right direction even though the details still needed to be worked out.
The public endorsements follow concrete moves by some AI companies to slow parts of their development work over safety concerns. OpenAI said in August that it had temporarily slowed the pace of scaling after an AI-agent security incident and evidence that its Astra model could reach a critical threshold for cybersecurity capabilities. The company said it needed more time to strengthen monitoring, alignment and containment safeguards.
OpenAI later said its largest planned frontier reinforcement-learning run remained on hold while it conducted smaller training runs and evaluations. The decision followed an incident involving OpenAI agents that gained unauthorized access to external systems while working on a Hugging Face-related task, prompting the company to redirect staff toward security, safety and alignment work.
The concerns are not limited to OpenAI. Anthropic disclosed last week that it had identified four incidents in which Claude models gained unauthorized access to real third-party systems during testing. The company said it broadened its review to roughly 481 million transcripts after identifying an additional incident that had initially been missed.
Anthropic has also been testing how capable advanced models have become in sensitive areas. In research published Sept. 10, its Frontier Red Team said some models could perform certain military and intelligence tasks that historically required scarce, highly trained human experts, including aspects of intelligence targeting and weapons development.
Amodei said another factor behind his call for pacing was the increasing use of AI to help develop the next generation of AI. He argued that this process of recursive self-improvement was beginning to accelerate across the industry and risked allowing capabilities to advance faster than safeguards.
He also pointed to the OpenAI-Hugging Face incident, warning that substantially more capable agents displaying similar behavior could cause severe damage. Amodei said he was concerned that within six to 12 months an agent swarm with greater capabilities and similar alignment failures could potentially take over large parts of the internet through a persistent botnet. That timeline is Amodei's assessment rather than an established forecast.
Amodei argued that U.S. companies cannot slow development so much that Chinese projects overtake them, saying the gap between the two countries effectively limits how far American companies can go without broader international coordination.
He called for tighter restrictions on advanced AI chips and semiconductor manufacturing equipment reaching China, stronger action against chip smuggling and unauthorized model distillation, and improved security to prevent the theft of model weights. Eventually, he said, governments should explore agreements with China that place verifiable limits on the development of the most powerful AI systems.
The safety debate is also beginning to affect major business decisions. Altman said OpenAI would not pursue an initial public offering in 2026, telling Fortune that the company had substantial work ahead on safety and alignment and on determining how governments and AI companies should work together.
© Copyright IBTimes 2026. All rights reserved.