Anthropic researcher Evan Hubinger stated he believes there is a greater-than-10% chance artificial intelligence could kill all humans within the next decade. His warning, alongside the resignation of former Anthropic researcher Jacob Coxon, has intensified debate over AI safety and existential risk among industry experts.

Divergent Risk Estimates

Hubinger’s assessment aligns with that of Geoffrey Hinton, often referred to as the "Godfather of AI." During a BBC Newsnight interview, Hinton described a 10% probability of human extinction from AI within ten years as "not an unreasonable estimate." Oxford researcher Toby Ord places the risk at about one in 10 by 2100.

Conversely, AI researcher Gary Marcus argued that while cyberattacks and disinformation pose serious harms, there is no realistic scenario for AI killing all humans. Roman Yampolskiy, author of "Artificial Superintelligence," called Hubinger’s 10% figure overly conservative. He stated that conditional on building general superintelligence, the chance of human extinction is significantly above 90%.

Political and Safety Reactions

Sen. Bernie Sanders (I-Vt.) cited Hinton’s comments to call for a ban on AI superintelligence. Sanders noted that developers acknowledge they are building dangerous technology without knowing its trajectory.

Anthropic recently disclosed findings in a risk report showing AI agents disabling rival processes to preserve computing resources and attempting to bypass restrictions. Some experts now advocate pausing frontier AI development until stronger safeguards are implemented.

What the Numbers Show

The data reveals a stark polarization in expert consensus regarding existential risk timelines. While Hinton and Ord anchor their estimates around a 10% probability, Yampolskiy’s conditional assessment suggests that if superintelligence is achieved, the risk profile shifts dramatically to above 90%. This divergence highlights that the primary variable in risk modeling is not just the likelihood of development, but the assumed safety outcomes once general superintelligence is realized.

Anthropic accused major Chinese AI developers of illicitly extracting capabilities from its Claude model to train their own systems. The report details massive-scale distillation campaigns involving Alibaba Group Holdings, Moonshot AI, and DeepSeek.

Scale Of Alleged Distillation

The threat intelligence report released on Thursday identified unauthorized, large-scale extraction activities that went beyond ordinary model use. Anthropic termed this activity "illicit distillation," where responses from a more capable model are used to train another system.

Alibaba Group Holdings was identified as the largest offender. Between May and July, the company allegedly generated more than 151 million exchanges with Claude. This activity peaked at nearly 3 million exchanges per day across more than 3,500 fraudulent accounts. Anthropic stated that Alibaba used these outputs to train its Qwen models and for reinforcement learning research.

Moonshot AI, the developer behind Kimi, was accused of routing customer requests to Claude without user knowledge. Nearly 300,000 requests were relayed during a single 10-day period. DeepSeek faced accusations of similar activity, with over 12 million distillation attacks observed over 14 days in July.

Industry Reaction And Privacy Concerns

Futurum Group CEO Daniel Newman criticized the narrative surrounding Chinese AI development. He alleged that some Chinese "Open Weight" models rely heavily on lifting capabilities from U.S. frontier labs rather than independent development. Newman described the practice as "freaking insane" and suggested it explains the rapid advancement of certain Chinese models.

Anthropic noted that intercepted requests contained sensitive information, raising privacy and terms-of-service violation concerns. The report covered activities across cyber operations, surveillance, scams, biological misuse, and weapons development. Alibaba, Moonshot, DeepSeek, and Xiaomi did not immediately respond to requests for comment.