Anthropic Researcher Warns of Over 10% Chance AI Could Wipe Out Humanity

Evan Hubinger, a safety researcher at Anthropic, said he believes there is more than a 10% chance that AI will exterminate humanity within 10 years. He wrote in a post on X that the danger from current foundation models is low. At the same time, he expressed concern that the technology could soon enter a stage of self-improvement, creating an existential danger. BeauHD’s reporting for Slashdot conveyed the remarks as covered by the BBC. The statement has been received not as an isolated view, but as one of a series of strong warnings emerging within the industry.

Researcher Puts Extinction Probability at Over 10%

Hubinger stated explicitly in a post on X that the danger from existing models is low. He then cited the extremely rapid pace of technological progress as the reason. He said the probability of an event leading to the death of all humanity within a 10-year period exceeds 10%. He did not detail the specific process of catastrophe. According to the Slashdot article citing BBC reporting, the point of his remarks lies not in assessing the present but in future change. The issue is not current uses or accidents, but controllability after a leap in capabilities. The post did not show the basis for calculating the probability or its underlying assumptions. Even so, it was taken seriously as a statement from someone responsible for safety research inside the company. The figure of 10% has sparked debate as a level requiring action, not a negligible margin of error. Attention was also drawn to the fact that the researcher himself described it as concerning.

The Core of Existential Risk from Self-Improvement

The core of his remarks is the shift to a stage where AI improves itself. Current models have a structure in which humans design, train, and deploy them. There is ongoing discussion that in the future, models could take on part of research and development, accelerating the cycle of improvement. Hubinger indicated his view that this transition is near. If self-improvement advances, understanding and verification of capabilities may fail to keep up. If deployment expands while evaluation remains incomplete, the impact of malfunctions or misuse will spread. By stating that current models pose low danger, he emphasized the difference in stages. The problem is not current response accuracy, but future autonomy and speed. The term existential danger is on a different level from accidents or unemployment. The scenario envisioned is one in which resources and infrastructure are seized in a way humans cannot intervene. However, no path to realization was shown, leaving much room for verification.

Former Researcher Challenges Corporate

Responsibility

The remarks were made in response to a post by Jacob Coxon, who left Anthropic. Coxon is a researcher with a history of working at OpenAI. In a post on X, he criticized both companies for not acting responsibly. In the original text, he stated as follows.

“Neither company is acting responsibly,”

He argued that systems arriving soon will reach a level beyond humans. He expressed the view that penetration of any target will become possible. He said they will gain the power to transform every field overnight. He indicated his view that they will go so far as to acquire real-world power and resources. Hubinger’s post was published in response to this argument. Warnings have thus come from both a researcher who left the company and one who remains inside, creating a striking structure. The remarks have drawn attention as whistleblower-like statements from parties to the development race. Questions are being raised about corporate governance and transparency of evaluation. The simultaneous outflow of talent and public criticism is also characteristic.

Risk Debate Shifts from Existence to Scale

BeauHD’s reporting for Slashdot presented an overview that the center of gravity of the controversy is changing. Previously, the very existence of danger was disputed. Now the dispute has shifted to how to estimate the scale of the danger. Hubinger’s remarks were positioned as the latest example of that trend. Presenting a figure of over 10% symbolizes the shift to debate over scale. Strong warnings continue to come from other researchers and executives. Discussion focused on existential danger is also spreading among regulators and academia. At the same time, criticism of warnings that show no concrete path remains strong. There is a view that priorities for countermeasures cannot be set while the basis for probabilities remains unclear. There is also a counterargument that exaggeration distorts investment and regulatory decisions. Even so, statements by those involved carry different weight from outside speculation. The fact that a sense from the development front has emerged as a numerical figure is significant.

Rising Tensions Over Export Controls and

Development Race

In the background are the intensifying development race and the search for regulation. How to manage the spread of high-performance foundation models is an international challenge. For discussion of export controls, see Anthropic Mythos Export Restrictions Risk Repeating PGP’s Mistakes for details. That article outlines the difficulty of technology diffusion and control. This warning also overlaps with the difficulty of control after diffusion. In a structure where capabilities are replicated across borders, restraint by a single company has limits. For the speed of the company’s foundation development, see Anthropic Declares Fable 5 Is Back. If new systems continue to be introduced, the burden of evaluation and operation will increase. Friction over the collection of training data is set against the background of conflict shown in Anna’s Archive Offers $200K Bounty for Full Google Books Scan. Rights clearance for data and the expansion of training remain unresolved. Ensuring safety requires not only computing resources but also institutional design. Frameworks for audits and incident reporting are being called for. The gap between the speed of development and the speed of verification underlies the danger debate.

Editorial Opinion

In the short term, screening of procurement and operations is expected to tighten. Companies will face more situations where disclosure of safety evaluations is demanded. Adopters will move to limit the scope of use and strengthen monitoring systems. In the long term, research and development of self-improving types and the design of verification systems are expected to become the focus. Frameworks for external audits and incident reporting may be established. Society is entering a stage of debating the distribution of the technology’s benefits and dangers as a whole. The basis for the 10% figure and how it is received are in question. The process of calculating the probability has not been disclosed. At what level of uncertainty should society change its behavior?

References

• “Anthropic Researcher Believes More Than 10% Chance AI ‘Could Kill All Humans’”, by BeauHD — Slashdot, 2026-09-09T21:00:00.000Z (ARR)
• Source URL: https://slashdot.org/story/26/09/09/1943258/anthropic-researcher-believes-more-than-10-chance-ai-could-kill-all-humans?utm_source=rss1.0mainlinkanon&utm_medium=feed