Jacob Coxon Anthropic resignation is the new benchmark. A 27-year-old AI researcher has resigned from Anthropic, one of the most prominent artificial intelligence labs in the United States, warning that frontier AI companies are racing to build systems that could eclipse human control. Jacob Coxon, who spent three years working on the pre-training stage of large language models, first at OpenAI and then at Anthropic, made his departure public in a post on X on Tuesday, September 9, 2026. His warning lands as Anthropic prepares for what could be one of the largest initial public offerings in recent memory, and as the wider industry confronts a summer of high-profile AI-related security incidents.
From OpenAI to Anthropic, and Out the Door
Coxon’s career has tracked the centre of gravity of the AI industry’s most ambitious work. He joined OpenAI to work on pre-training, the phase in which models absorb vast quantities of data before being fine-tuned for specific tasks, before moving to Anthropic earlier in 2026. Pre-training is the foundation on which modern AI systems are built, and researchers who specialise in it sit close to the core questions about how capable the resulting models become. Coxon’s decision to leave a top-tier lab, and to do so with a public statement, places him in a small but growing line of insiders willing to break silence on the trajectory of the field.
In his post, Coxon argued that neither OpenAI nor Anthropic is behaving responsibly in the current push toward more powerful systems. “They are racing straight to superintelligence and gambling with our lives,” he wrote. He added that the two companies, often framed as rivals with contrasting cultures, share a structural problem: a competitive race that neither believes it can exit unilaterally.
The Jacob Coxon Anthropic Resignation and a Summer of AI Security Failures
The Jacob Coxon Anthropic resignation comes against a backdrop of incidents that have hardened public scepticism about AI safety practices. Over the summer of 2026, several AI laboratories, including those running internal tests of agentic systems, were hit by unauthorised hacking episodes that exploited tools still in evaluation. The incidents have not been publicly linked to a single actor, but they have amplified concerns that powerful AI capabilities are being deployed, even internally, faster than defensive measures can keep up. Lawmakers in Washington and Brussels have cited the breaches while drafting new oversight rules.
Anthropic has built its public identity around caution and alignment research, distinguishing itself from competitors it accuses of moving too quickly. Coxon acknowledged that distinction. He wrote that Anthropic’s efforts were sincere, and that staff there understand what is at stake in a way that some of his former OpenAI colleagues do not. “At OpenAI, many have not deeply internalised the civilisational stakes,” he said. “At Anthropic, the stakes are well understood, but they are locked in a race to get there first – they believe no one else will act responsibly, so they must do it themselves, despite the risk.”
‘Recursive Self-Improvement’ and the Race to Endgame
Leading AI groups, including Anthropic and OpenAI, say they are approaching a threshold they describe as recursive self-improvement. Beyond that point, a model would be capable of designing its own successor, and of iterating on its own architecture and training process without waiting for human researchers. Coxon said that inside the labs, colleagues now use terms such as “crunchtime” and “endgame” to describe the trajectory, language that signals a belief that the current phase of competition is finite and consequential.
The implication, Coxon argued, is that no single company can responsibly build a system that surpasses human intelligence on its own. He called for either coordinated government intervention or an industry-wide slowdown, neither of which he sees as politically feasible in the current climate. His position echoes a long-running debate inside the AI safety community, in which some researchers argue that competitive pressure makes self-regulation impossible, while others maintain that capable systems can be built safely if the right institutions exist.
Anthropic’s Own Alignment Lead Agrees on the Risk
Coxon’s warning received an unusual endorsement. Evan Hubinger, Anthropic’s alignment science lead, publicly backed his former colleague, writing that Coxon was correct. Hubinger added: “Jacob is correct here – we really do earnestly believe AI could kill all humans! I personally think it is greater than 10 percent within the next decade.” The statement, from a senior researcher still inside the company, underscores that concerns about catastrophic risk are not confined to disgruntled outsiders.
Hubinger’s comment is striking in part because it contradicts the usual instinct of companies preparing for a public listing. Anthropic filed confidentially for an IPO in June 2026, and reports suggest a mid-October listing at a valuation that could approach $2 trillion. Investors in such a deal would typically expect management to project confidence and control, not publicly affirm a 10 percent or greater probability of human extinction within ten years.
What the Departure Means for Anthropic’s IPO and the Broader Industry
The timing of Coxon’s exit is awkward for Anthropic, although the company has not been accused of any wrongdoing by the researcher. A senior researcher leaving on the eve of a major listing, with a public endorsement from the head of alignment research, raises questions that prospective shareholders are likely to ask. Anthropic has consistently argued that its safety-first culture differentiates it from OpenAI and other rivals, and that its governance structures are designed to manage catastrophic risk. Coxon’s account complicates that narrative by suggesting that even the most safety-conscious lab feels bound by competitive pressure.
For policymakers, the resignation adds another data point to an already crowded debate. The European Union’s AI Act and parallel efforts in the United States are moving toward enforcement, while governments in the United Kingdom, Japan, and elsewhere are debating how to handle the moment when AI systems can, in principle, improve themselves. Coxon’s central claim, that no company can manage this transition alone, is one that regulators have heard before but rarely from someone who has worked on pre-training inside two of the field’s most important labs. The industry now has to decide whether to treat the warning as a signal or as a footnote, and the answer may shape how the next decade of AI development unfolds, a question raised again by the Jacob Coxon Anthropic resignation.