Three Anthropic Researchers Warn of Real Risk of Human Extinction Within a Decade
Three leading researchers from Anthropic publicly warned that frontier labs are racing towards potentially uncontrollable superintelligence, believing there is a real risk of human extinction within this decade.
Pre-training researcher Jacob Coxon announced his resignation, stating that during his three years at OpenAI and Anthropic, both companies acted irresponsibly, racing towards self-improving superintelligence. He told The Wall Street Journal that in a more aggressive scenario, it could be out of control by the end of next year.
Evan Hubinger, head of alignment science who is still employed, publicly agreed, stating that colleagues indeed believe AI could kill everyone, and he personally estimates the probability of this happening in the next decade to be over 10%. He distinguished that the risks of existing models are still low, but the real concern is superintelligence arising from recursive self-improvement, which is progressing faster than previously judged.
Safety researcher Samuel Marks spoke out personally, stating that Coxon's narrative is worth reading; developers believe the technology could lead to extinction or equally severe consequences, and the higher the rank, the greater the concern. Hubinger also clarified that the company is making efforts, but there is still no solution for superintelligence alignment, nor can it be said that they are on that track.
Coxon described OpenAI as not having internalized the stakes of civilization, and Anthropic as having recognized the risks but rushing into the endgame because they believe their competitors will not stop. He advocates for U.S. labs to establish a pacing agreement, suggesting that the global race may need to temporarily halt further enhancements of model capabilities.
In market mechanisms, this is a reputational and talent event: the safety premium is rewritten by internal witnesses, while training and financing schedules are still priced by the computational power race. Beneficiaries are regulatory discussions, training halting initiatives, and risk hedging narratives; those under pressure are Anthropic, which relies on the "more responsible lab" brand for financing and hiring, as well as peers' safety communication standards. Funding will not stop building clusters due to posts, but insurance, policy, and corporate procurement will reprice extinction probabilities.
The statements from the three coincide with the earlier departure of Anthropic's head of safety and the admission by OpenAI's chief scientist that alignment has not been resolved, turning private fears into publicly quotable probabilities.
Source: Public Information
ABAB AI Insight
Anthropic relies on the safety faction of OpenAI, now being simultaneously testified by its own pre-training, alignment, and safety positions: risks are acknowledged, but solutions are not. Coxon provides capability-side testimony, Hubinger provides probabilities within the official alignment framework, and Marks indicates that higher ranks correlate with greater fear. These three points form a chain: the more one knows, the less they believe the race can internally brake.
Capital continues to invest according to model generations and cloud contracts, with alignment budgets as a supplementary item. Hubinger's statement that "there is no superintelligence alignment plan and we are not on track" effectively shifts the safety department from braking to a narrative role. Labs are still increasing pre-training because stopping would mean handing over the rights to define the next generation of systems to competitors. Resources are locked in computational power and release windows; testimonies can only change public opinion, not training queues.
In contrast to Sharma's departure, Leike leaving OpenAI, and Turner leaving DeepMind: exits from safety roles or public dissent have become a fixed outlet in the industry. The phase has shifted from expanding the safety brand to self-refutation: externally still selling responsibility, internally acknowledging that the endgame is driven by private company schedules.
This reflects a governance vacuum before technological replacement is complete: extinction probabilities are internally priced, and decision-making power remains within the competitive structure. The mechanism is that acknowledging risks does not generate veto power; veto power only exists for those who first achieve self-improvement.
ABAB News · Cognitive Laws
Those who acknowledge risks do not necessarily hold the brakes.
The higher the rank, the greater the fear, indicating more complete information at the upper levels.
A race without an alignment plan means the plan itself is to continue racing.