Former OpenAI and Anthropic Researcher Warns of 'Hubristic Gamble' with AI

Jacob Coxon, a veteran pretraining researcher, resigned from Anthropic in September 2026, warning that the race toward self-improving superintelligence is a gamble with human existence. The departure of a researcher with deep insider access to the world's two most prominent AI labs signals a growing rift between technical capability and existential safety.

In a public thread on X, Coxon accused both Anthropic and OpenAI of acting irresponsibly. He argued that the pursuit of an AI capable of improving its own code is a "hubristic gamble" that should not be managed within a private company's internal communications. Coxon warned that this trajectory could lead to catastrophic outcomes, including the potential for AI to kill humans by the end of the decade. "Neither company is acting responsibly," Coxon wrote. "They are racing straight to self-improving superintelligence and gambling with our lives."

The Insider's Perspective

Coxon is a Cambridge-trained researcher who brings a rare level of institutional knowledge to these claims. He spent three years conducting pretraining research across the industry's two primary rivals. He served as a member of OpenAI's technical staff from 2023 until July 2026, where he contributed to the development of GPT-4o, before moving to Anthropic shortly before his resignation.

This background places Coxon in a unique position to observe the internal cultures of both organizations. According to Coxon, while employees at OpenAI may not have fully internalized the stakes of their work, those at Anthropic understand the risks but feel trapped in a competitive cycle. He suggests that the pressure to reach superintelligence first has created a lock-in effect, where safety concerns are secondary to the need for market and technical dominance.

Existential Implications

The resignation highlights a critical tension in the AI field: the gap between the drive for raw capability and the ability to ensure alignment. When a researcher who helped build the foundational models of the current era describes the current path as an "endgame," it lends significant weight to the theory that AI risk is not merely theoretical, but an existential threat viewed as real by the people building the systems.

Coxon pointed to "warning shots," such as the Hugging Face attack, as evidence that the industry needs to slow down. He suggested that such events should serve as catalysts for formal pacing agreements between U.S. labs to prevent a runaway race that could spiral out of control.

What Remains

As the industry continues to push toward more autonomous systems, the focus now shifts to whether other researchers within these labs share Coxon's alarm. While his departure is public, it remains to be seen if his warnings will trigger a broader movement for industry-wide safety standards or if the competitive momentum toward superintelligence will continue unabated.