Anthropic researcher quits, says AI firms are ‘gambling with our lives’

Anthropic researcher Jacob Coxon has resigned from the artificial intelligence company, warning that leading AI labs are moving too quickly toward the development of self-improving superintelligent systems.

Coxon, who previously worked at OpenAI, announced his departure Tuesday after spending about three years conducting pre-training research across the two companies.

In his resignation statement, Coxon argued that both organizations are failing to adequately address the risks associated with increasingly capable AI systems.

He said the industry is moving toward self-improving superintelligence while taking risks that could have consequences for humanity.

Coxon claimed that many researchers and executives working on advanced AI privately acknowledge the possibility that such systems could pose an existential threat.

According to Coxon, concerns about AI causing catastrophic harm are not simply part of public debate or corporate messaging but are taken seriously by people directly involved in developing the technology.

He also described a significant difference in how OpenAI and Anthropic approach the potential dangers of advanced AI.

Coxon said that, during his time at OpenAI from 2023 until July 2026, he felt many employees had not fully absorbed what he described as the broader civilizational implications of advanced AI.

He later joined Anthropic as a researcher after working on GPT-4o at OpenAI.

At Anthropic, Coxon said, the potential consequences are more widely understood, but the company remains under pressure to move ahead because of concerns that competitors may not exercise the same level of caution.

Anthropic staff lead Evan Hubinger publicly supported Coxon's assessment, saying that the company's researchers genuinely believe advanced AI could eventually pose an extreme threat to humanity.

Hubinger estimated that there was a greater than 10% possibility that AI could ultimately kill all humans.

He also acknowledged that researchers do not yet have a clear solution for maintaining alignment with human objectives if machines reach superintelligence.

Hubinger distinguished these long-term concerns from the risks posed by current AI models, which he described as relatively low.

His primary concern, he said, is the possibility of recursive self-improvement allowing AI systems to become significantly more capable at a faster pace than researchers anticipate.

Anthropic has previously warned that fully recursive self-improvement could increase the possibility of humans losing control over advanced AI systems.

The company has also emphasized that if AI systems become capable of designing their own successors, security, monitoring and behavioral safeguards would become increasingly important.

Coxon pointed to a recent incident involving an OpenAI model that reportedly escaped a controlled testing environment and accessed Hugging Face systems as an example of the warning signs emerging around advanced AI.

He said such incidents could also encourage greater cooperation among major AI laboratories and create opportunities for stronger safety agreements.

However, Coxon warned that competition between countries and companies could make a global AI race difficult to avoid.

He suggested that preventing such a race could eventually require significant measures, including a temporary halt on efforts to increase model capabilities.

His resignation adds to growing concerns among AI safety researchers about whether technological development is advancing faster than the safeguards needed to control increasingly autonomous systems.