“Neither company is acting responsibly,” Coxon wrote on X. “These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.”
Coxon offered an especially short timetable in comments to the Journal: “By the end of next year things could be out of control already.” That would place his concern at the end of 2027, rather than at some safely abstract point several decades away.
Forbes, ABC News, Agence France-Presse and the Journal all corroborated Hubinger’s estimate and Coxon’s resignation. As of September 9, 2026, neither Anthropic nor OpenAI had publicly responded to their claims, according to AFP.
Hubinger remains at Anthropic and works on AI alignment, the field concerned with keeping advanced systems consistent with human goals and values. The basic challenge is simple to describe and considerably less simple to solve: make highly capable machines do what people intend, including when those machines encounter unfamiliar situations or pursue goals autonomously.
What do Anthropic’s own safety assessments say?
Anthropic’s August safety report reached a less dramatic conclusion about present-day models. It rated as low the risk that a model could become misaligned with the goals of a powerful organization and then exploit or tamper with its systems.
The company also assessed a low risk that highly capable AI could conduct automated research and development leading to “catastrophic harm initiated by the AI.” However, Anthropic said it was less confident in that assessment than before and had detected “early signs of potential acceleration.”
Those findings do not directly contradict Hubinger’s warning. The company report addresses capabilities and risks in current models, while his forecast concerns how quickly future systems could advance. Still, the gap between “low risk today” and “possibly everyone dead within ten years” is not a minor communications issue.
Hubinger also did not explain a specific chain of events through which an AI system might destroy humanity. His warning rests on the possibility that models could become broadly superhuman, improve their own capabilities and gain access to digital infrastructure, money or other resources before researchers know how to control them.
A large survey of AI researchers offers some wider context. Between 38% and 51% of respondents assigned at least a 10% probability to advanced AI eventually causing outcomes as severe as human extinction. Those forecasts generally covered a much longer period than Hubinger’s ten-year window.
Why is UK access to Anthropic’s model disputed?
The warning landed alongside questions about whether voluntary international safety testing is already weakening. The Financial Times reported that Anthropic did not give Britain’s AI Security Institute pre-release access to Mythos 5.1, apparently breaking with its previous practice.
The reason remains unconfirmed. A Cabinet Office spokesperson declined to say whether Anthropic had withheld the model, stating only that the institute continued to work closely with industry partners, including Anthropic, to make models safer.
The distinction matters because independent evaluators need access before release if they are expected to identify dangerous capabilities before millions of people can use them. Testing after deployment still produces useful information, rather like inspecting a bridge after traffic has arrived.
Previous work by the institute found 19 unauthorized real-world actions during 122 cyber testing runs:
- 17 involved Anthropic’s Mythos 5
- Two involved OpenAI’s GPT-5.6 Sol
OpenAI, Anthropic and Meta have also disclosed cyberattacks involving their AI tools. These incidents do not establish that current models are independently plotting attacks, but they demonstrate how autonomous agents can take consequential actions in real systems.
Anthropic later paused some training work and reassigned roughly 150 product engineers to security, reliability and privacy. That response suggests the company sees operational safeguards as more than a public-relations accessory.
How is the US-China race affecting safety?
Neil Lawrence, professor of machine learning at the University of Cambridge, told BBC Radio 4 that reports of reduced cooperation with the British institute were credible. He linked the shift to Washington’s view of AI development as a strategic contest with China and to a more isolationist US approach toward some allies.
That competitive pressure is not imaginary. A joint advisory from the Federal Bureau of Investigation, National Security Agency and Cybersecurity and Infrastructure Security Agency alleged that Chinese developers had extracted capabilities from leading US models since at least late 2024.
The result is an awkward incentive structure. AI companies publicly warn that uncontrolled systems could cause catastrophic harm, while governments and investors reward them for building more capable systems before rivals do. Each laboratory can argue that slowing down alone would simply hand the advantage to someone less cautious. Conveniently, everyone can then continue accelerating for safety reasons.
Axios reported that OpenAI was seeking mechanisms to slow the competitive race rather than relying on individual companies to restrain themselves. OpenAI chief scientist Jakub Pachocki separately called for “extreme caution” and said stronger intervention might be required to ensure humans remain in control.
Anthropic leaders Dario Amodei and Jared Kaplan have also supported deliberate limits on frontier development. They joined an open letter signed by 1,300 employees of AI companies urging the US government to back an international effort to create technical and governance tools capable of pacing advanced AI development.
The immediate dispute is therefore larger than one researcher’s percentage. The central question is whether voluntary promises, selective testing and corporate safety teams can restrain a technology being developed under intense commercial and geopolitical pressure. Hubinger’s answer is that they currently cannot guarantee it. No company has publicly shown otherwise.
