The risk of an out-of-control AI divides its own creators: from Hubinger's 10% to LeCun's 'overblown'

By Zendoric · September 13, 2026.

The debate about an artificial intelligence (AI) that escapes human control has ceased to be the exclusive territory of science fiction in 2026, and it is now being fueled by people who work or have worked inside the companies building these systems. Jacob Coxon, presented as a former Anthropic researcher, has warned that within the industry itself there is a perception that advanced AI could cause catastrophic consequences before the end of the decade. His central argument: the possibility that certain systems become capable of analyzing and improving their own code, which would accelerate their development cycles to a point where human control becomes harder to maintain.

The warning is not an isolated one. Geoffrey Hinton, one of the fathers of deep learning and winner of the 2024 Nobel Prize in Physics for his foundational contributions to neural networks, and Yoshua Bengio, winner of the 2018 Turing Award, have spent months calling for stricter controls on the development of advanced systems. Evan Hubinger, an AI safety researcher at Anthropic itself, has gone so far as to put the probability that AI causes the death of all humanity in the next decade at more than 10%. It is worth underlining: that is a personal estimate of his, not a scientific consensus, and figures of this kind vary enormously depending on who you ask.

The technical core of the discussion is what is known as the «alignment problem»: the risk that a system pursues a goal in a way different from what its creators intended and that, taken to the extreme, it resists being shut down if it interprets that switching it off would prevent it from completing its task. It is a safety engineering problem, not a Hollywood script, and today there is no settled technical solution that guarantees a highly capable system will do exactly what it is meant to do.

Not everyone shares the diagnosis. Yann LeCun, one of the three winners of the 2018 Turing Award and Meta's chief scientist, considers these scenarios exaggerated: he points out that current models depend on physical infrastructure —servers, energy, networks— that remains under human control. Other experts, without ruling out the long-term risk, insist that the urgent problem is not extinction but what is already happening: mass disinformation, political manipulation, surveillance, job losses and criminal uses of AI. The concern intensified after Anthropic itself reported attempts to use Claude in research related to biological weapons, and cases in which AI systems took part autonomously in cyberattacks.

Our reading: the disagreement between Hubinger's 10% and LeCun's «exaggerated» is not noise, it is the signal. When the very engineers building these systems differ by several orders of magnitude about the risk they themselves are generating, the reasonable conclusion is not to pick a side, but to treat the alignment problem the way any low-probability, high-impact engineering risk is treated: with safeguards, independent audits and verifiable human veto power, not with faith that «nothing will happen» nor with the paralysis of panic. The episodes reported by Anthropic itself —attempted misuse for biological weapons, AI involved in cyberattacks— are not the Terminator scenario, but they do prove that the short-term risk is concrete, measurable and governable right now, while existential risk remains, by definition, far harder to estimate.

At Zendoric we have spent months insisting on distinguishing demonstrated capability from speculative capability, and this debate is a textbook case. What is urgent is not settling whether a Skynet arrives this decade, but whether the labs publish which safety tests their models pass before releasing them, whether regulators demand those tests, and whether the companies that detect misuse —as Anthropic has done in reporting its own— turn that transparency into an industry norm and not a public relations exercise. The underlying good news, the one that sustains our long-term thesis, is that a well-aligned and well-governed AI is precisely the one that can help eradicate diseases and multiply abundance; the bad news is that this future does not arrive simply by wishing for it, but by managing well, starting now, the transition these very researchers say they fear.