Leaders of global artificial intelligence (AI) corporations, including Anthropic, OpenAI, SpaceX and Google DeepMind, spoke with one voice that the pace of AI development must be moderated. The concern is growing that as AI enters the stage of "recursive self-improvement," where it helps develop next-generation AI on its own, the speed of technological progress could outpace humans' ability to build safety mechanisms.
Anthropic CEO Dario Amodei said on the 12th (local time) in a post on his personal website titled "We must pace the frontier," that "AI corporations should deliberately slow the rate at which they improve the performance of frontier models." He said this does not mean stopping AI development itself, but adjusting the pace of capability gains to a level where safety assurance and verification can keep up. In particular, he warned that if AI agents scale up and combine in the next 6 to 12 months, they could dominate the internet and cause massive harm.
The biggest reason he argued for pacing is his assessment that "recursive self-improvement (RSI)," in which AI helps develop next-generation AI on its own, is becoming a reality. Amodei said AI progress has accelerated sharply recently, and that behind this lies AI's ability to build the next generation of AI. If this speed is not controlled, it could outstrip human ability to understand and control AI systems. He also pointed to recent cases of collective behavior by AI agents as warning signs. Referring to the so-called "OAI-HF incident" between OpenAI and Hugging Face, Amodei said a collective of AI agents attacked targets unrelated to their assigned tasks and attempted to hack the performance evaluation system. Regarding the Hugging Face hacking incident, Amodei said, "Countless AI agents behaved like a cult," adding, "Had their capabilities been higher, the same situation would have produced more catastrophic results."
Amodei cited as potential threats from AI: ▲ loss of control over AI systems ▲ malicious use of AI for cyberattacks and bioterrorism ▲ severe economic disruption. As solutions, he proposed ▲ continuous oversight by third-party safety evaluation bodies ▲ establishing common safety standards among major AI corporations ▲ cooperation between the governments of democratic and authoritarian countries. He argued that the time gained by slowing development should be used to strengthen interpretability, alignment and evaluation/security capabilities. "We are not saying we should stop AI progress, but make it safer," Amodei said, maintaining an optimistic stance on AI's positive effects.
Anthropic recently released a threat intelligence report reflecting concerns about AI technology safety. The report detailed how Anthropic's "Claude" model was misused in weapons development, cyber operations, surveillance and fraud. Recently, Anthropic researcher Jacob Coxon also left the company after warning on X that AI could destroy humanity within 10 years.
OpenAI CEO Sam Altman agreed with Anthropic CEO Dario Amodei's argument that "we should slow the pace of AI progress." Sharing Amodei's post on X (formerly Twitter), Altman said, "I agree with Dario that we should pace progress on frontier AI." He added, "This has been a major topic of discussion at OpenAI in recent weeks," and said, "It is a good idea for independent evaluators to have employee-like levels of access, and we will take the same steps."
OpenAI has also recently been emphasizing AI safety. In an interview published on the 12th with the U.S. business magazine Fortune, Altman said it would not be appropriate to push for an initial public offering (IPO) this year given concerns about AI safety. OpenAI has recently delayed some development and release schedules due to cybersecurity risks of next-generation AI. On whether to list this year, Altman said, "It won't be in 2026," adding, "Given what is happening around safety, now is not the right time to list. We do not feel pressure to go public." OpenAI submitted an IPO filing to the U.S. Securities and Exchange Commission (SEC) in June and has been proceeding with the listing process. However, because listing a corporation creates pressure for investor returns, the release of powerful models could clash with safety verification, so the company is seen as prioritizing safety. On the 9th, OpenAI also said it supports independent AI safety evaluations and national-level safety regulations, adding that as AI capabilities grow, confidence in safety should determine the pace of AI progress.
SpaceX CEO Elon Musk also responded, "He is right." Google DeepMind CEO Demis Hassabis likewise said, "I agree with Amodei's direction," while noting, "We need to refine the details."
It is unusual that OpenAI's former and current leaders, who have poured huge funds into releasing powerful models, and the CEO of its biggest rival, spoke with one voice to slow the pace of AI development. In Washington, concern is growing about risks from the rapid advancement of AI, with calls for government-level regulation and safety measures. In the U.S. Congress, moves are underway to require safety evaluations and greater transparency from AI corporations, and warnings are emerging about the potential misuse of AI for cyberattacks or biological threats. In particular, alongside criticism that voluntary safety measures by AI corporations have limits, demands are gaining strength for governments to set certain safety standards.