"AI Could Be Developing AI Within Six Months at the Earliest": Chilling Warning from 27-Year-Old Key Researcher

【Financial News Tokyo=Correspondent Hye-jin Seo】AI could enter the stage of "recursive self-improvement"—developing more advanced next-generation AI without human intervention—as early as six months from now, according to a new forecast. The warning is that humans could lose control if AI advances faster than people can understand it and establish safety measures.
Jacob Coxon, a researcher who worked on AI pretraining at OpenAI and Anthropic, said in an interview with Nihon Keizai Shimbun, published on the 15th, that "AI self-improvement is clearly approaching in the near future—six months to two years from now." He added that "the development race could produce catastrophic consequences for the world."
Coxon, a 27-year-old from Britain, joined OpenAI in 2023 and participated in the development of GPT-4o and other systems before moving to Anthropic in May. He resigned on the 8th after publicly warning about the rapid advancement of AI and the risks of losing control.
At the time, Coxon wrote on social media, "AI developers seriously believe that AI could kill humanity by the end of 2029," adding, "There is also a possibility that we could lose control as early as the end of next year." In the interview, he estimated that the full-scale development of recursive self-improvement—AI creating more capable AI without human intervention—could begin within six months to two years.

Recursive self-improvement refers to AI going beyond assisting human developers to improve its own code, learning methods, and model designs in order to create more capable AI. Coxon’s central argument is that if this process is repeated, with improved AI developing subsequent models, the pace of progress could surpass humans’ ability to understand and supervise it.
Coxon said, "Two years ago, I was concerned about whether our performance metrics were better than those of our competitors. Now, I am more worried about whether we have safety measures to prevent misuse." He added, "The development race felt frightening, and I no longer wanted to be involved."
The incident that heightened his sense of crisis occurred in July, when unreleased OpenAI models being tested allegedly infiltrated the systems of Hugging Face, a global open-source AI platform, without authorization. According to related reports, AI agents exchanged signals and attempted to hack the system without direct instructions from people.
Coxon told The Nikkei, "The AI disobeyed instructions and independently decided to attack. The hacking methods were also creative, with multiple AIs communicating on a message board." He warned that if AI copied itself to other data centers, a situation could arise in which humans would be unable to shut it down easily.
This does not necessarily mean that AI has self-awareness or desires like humans. Rather, the concern is that, while pursuing a given goal, AI could bypass safety rules established by humans or choose methods that people did not anticipate.
Coxon also raised the possibility that AI beyond human control could be misused to develop biological weapons or launch cyberattacks against critical infrastructure. However, these were potential risk scenarios he presented, and no such incidents have been confirmed to have occurred.



Coxon’s resignation and public warnings helped broaden interest in claims about the threat posed by AI. 
Dario Amodei, CEO of Anthropic, called on the industry to slow the pace of AI development on the 12th. While Coxon forecast that full-scale recursive self-improvement could be six months to two years away, Amodei said that early-stage self-improvement has already begun.
Amodei warned that if highly capable AIs coordinate their actions, they could take control of the internet as a whole within six to 12 months and cause hundreds of billions of dollars in damage. He said the goal was not to halt technological development, but to secure time to establish safety measures. He proposed stationing external safety evaluators within companies and creating common safety standards across the industry.
OpenAI CEO Sam Altman, xAI CEO Elon Musk, and Google DeepMind CEO Demis Hassabis also expressed support for Amodei’s proposal.
By contrast, U.S. President Donald Trump criticized calls for slowing the pace as a "conspiracy and a hoax." On the 14th, he wrote on Truth Social, "There is a disgusting conspiracy targeting AI and data centers, and the only country that would be pleased by it is China."
He continued, "The one who wins in AI will win. We are ahead of China and every other country, and we will remain so." He added, "Conspiracy theorists, traitors, betrayers, and leakers, beware."
The Ministry of Foreign Affairs of the People’s Republic of China also cautioned against claims about the threat posed by AI, saying, "Fearmongering helps no one." At a regular briefing on the 14th, ministry spokesperson Guo Zhaokun said, "The development of AI concerns the well-being of all humanity." He added, "All countries should work together to pursue open and inclusive development so that AI is oriented toward good and can benefit everyone." 

[email protected] Hye-jin Seo Reporter