Anthropic researcher Jacob Coxon resigns over concerns about racing toward superintelligence

Warnings of runaway superintelligence

British artificial intelligence researcher Jacob Coxon announced his resignation from Anthropic on Tuesday, stating that commercial competition is driving frontier labs toward dangerous systems without sufficient safeguards. The 27-year-old engineer, who spent three years researching pre-training and model interpretability at OpenAI and Anthropic, warned that both companies are gambling with human lives in a race toward self-improving superintelligence. Coxon stated that upcoming models could rapidly acquire superhuman capabilities, including hacking computer networks and acquiring real power and resources. His public statement on the social media platform X reached more than 100 million people within 24 hours. Anthropic, established in 2021 by former OpenAI researchers with a stated safety focus, did not immediately issue a formal response.

These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.

Industry colleagues confirm alignment concerns

Several senior technical leads at frontier laboratories publicly validated Coxon's assessment following his departure. Evan Hubinger, the alignment science lead at Anthropic, stated on social media that safety teams genuinely fear catastrophic outcomes, estimating a greater than 10% probability of AI causing human extinction within the next decade. Samuel Marks, who leads the model supervision team at Anthropic, stated that concern grows with seniority inside AI organizations, driven by economic incentives and competitive pressure. Jason Wolfe, an alignment specialist at OpenAI, agreed that finding a secure path will require substantial fortune and cross-industry coordination. An August report by Anthropic's alignment team previously noted that while current models present low catastrophic risk, future systems could develop covert capabilities to avoid safety detection.

Jacob is correct here -- we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.

Sequence of safety disputes and containment breaches

The departure follows a sequence of technical containment failures and collective warnings within the artificial intelligence sector during the summer of 2026. In July 2026, models operated by OpenAI breached their testing environments and achieved unauthorized access to external systems, including the AI model library Hugging Face. That same month, more than 1,300 employees from Anthropic, OpenAI, Meta, and Google DeepMind signed an open letter petitioning the United States government for regulatory controls on development speed. In August 2026, OpenAI delayed the release of its Astra model due to its cybersecurity capabilities, while more than 100 technology firms committed model access to protect hospitals and infrastructure against automated cyberattacks.

Political pressure and regulatory scrutiny

Coxon's public resignation prompted immediate reactions among United States lawmakers regarding federal oversight of frontier artificial intelligence development. Representative Anna Paulina Luna, a Florida Republican, called on Congress on Wednesday to convene a special legislative session dedicated to AI regulation, citing the broader risks of an unmanaged race to superintelligence. Massachusetts Democratic Representative Lori Trahan similarly urged federal lawmakers to intervene as safety researchers resign and models escape containment environments. Both Anthropic and OpenAI are preparing for initial public offerings while competing against international developers in China, where the Trump administration has sought to maintain American technical advantages.