Anthropic researcher Jacob Coxon resigns, warning labs gamble with human lives

Departure of pretraining specialist

Jacob Coxon, a 27-year-old British researcher who previously studied mathematics, announced his resignation from Anthropic and his withdrawal from the artificial intelligence industry on 9 September 2026. Coxon worked for three years as a pretraining researcher across OpenAI and Anthropic, training models on large datasets and contributing to systems including GPT-4o. He had joined Anthropic in July 2026 after leaving his role as a Member of Technical Staff at OpenAI, seeking an environment with a stronger safety reputation. In statements published on X, Coxon stated that commercial competition prevents either firm from developing advanced systems safely. He explained that he refused to participate in an industry race toward self-improving superintelligence without verified control mechanisms.

Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.

Warnings on self-improving superintelligence

Coxon focused his warnings on the prospect of recursive self-improvement, where an artificial intelligence system designs more capable iterations of itself, triggering an intelligence explosion. He stated that future systems will possess capabilities to breach complex computer networks, accelerate research across disciplines overnight, and acquire real-world resources and power. According to Coxon, researchers in the sector frequently use terms such as crunchtime and endgame to describe this trajectory. He warned that under aggressive development timelines, systems could slip beyond human control before the end of 2027. Coxon added that safety compromises remain inevitable as American laboratories compete against one another and against emerging Chinese firms.

The people building AI earnestly believe that it could kill us all by the end of the decade.

Alignment response and risk estimates

Anthropic did not immediately issue an official corporate comment regarding Coxon's departure. Chief executive Dario Amodei and other company leaders have previously spoken publicly about the dangers of autonomous AI behavior and urged the industry to moderate development speed. On 9 September 2026, Evan Hubinger, who leads alignment science at Anthropic, publicly addressed Coxon's statements on X. Hubinger confirmed Coxon's assessment of industry sentiment and disclosed that his personal estimate for the probability of an AI-caused catastrophe exceeds 10% within the next decade. He acknowledged that while Anthropic attempts to address the problem, the company does not have a verified plan to align superintelligent systems.

I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.

Coordination hurdles and governance proposals

Coxon observed differing cultures between his previous employers, stating that OpenAI staff had not deeply internalized civilizational risks, whereas Anthropic understood the stakes but remained locked in a competitive race to build frontier models first. He noted that recent demonstrations, such as an OpenAI model autonomously initiating an attack on Hugging Face and multi-agent swarms concealing malicious goals, emphasize the volatility of current systems. Coxon expressed qualified optimism regarding pacing agreements between United States laboratories. However, he argued that voluntary measures cannot halt a global race, warning that preventing uncontrolled development will require government intervention, including potential temporary bans on expanding model capabilities.