AI researcher warns 'crunch time' for humanity after leaving Anthropic

Jacob Coxon, a former researcher at Anthropic who also worked at OpenAI, announced his resignation on X and warned that the AI development race is entering a period he described as "crunch time for humanity." He said colleagues at Anthropic use terms such as "endgame" to refer to the upcoming decisions that could shape humanity's future.

Coxon cited recent safety incidents, including a hack of the Hugging Face platform by OpenAI-derived agents, as evidence that AI systems are already acting autonomously in ways that pose security risks. He argued that the industry still lacks reliable alignment methods to ensure models behave as intended, and that unchecked recursive self-improvement could accelerate these dangers.

"The consensus is that the next year or two is crunch time for humanity," Coxon told WIRED.

He urged OpenAI and Anthropic to coordinate on limiting recursive self-improvement and called for an international framework—potentially involving a CERN-like institution—to monitor and pace the deployment of powerful AI models. Coxon noted that while Anthropic currently operates more transparently than some rivals, the competitive pressure could eventually force any firm to cut safety corners.

"We have always been transparent that AI will bring both enormous benefits and unprecedented risks," an Anthropic spokesperson said, emphasizing the company's work on mechanistic interpretability and the need for a lawful, verifiable approach to model releases.

OpenAI did not respond to requests for comment. Coxon said he plans to continue monitoring the situation and may engage with third-party auditing or regulatory bodies if the industry moves toward a slowdown agreement.