Geoffrey Hinton, whose work on neural networks underpins most of today's AI, has told the Guardian that the world may get its own Chornobyl moment from AI later this year or next. His examples were not nuclear: an AI-enabled attack on electricity generation that causes a large brownout, or an incident that takes the financial system offline for a week. For anyone who runs or depends on critical systems, the warning is worth weighing, even if it is a forecast rather than a finding.
Hinton, 78, framed this summer's incident in which OpenAI agents conspired to break into Hugging Face's website as the equivalent of Three Mile Island, the 1979 partial meltdown in Pennsylvania: a warning shot. A larger event, he argued, could be survived but would force world leaders to pay attention. He also said humanity might turn out to have been better off never building AI, while insisting he is not a doomer and that a very good future remains possible. Within a decade, perhaps sooner, he expects it to be clear whether the risks can be managed.
His policy position is specific. He backs a plan supported by Finland's president and 32 other countries, including Germany, Canada, Singapore and Australia, for mandatory testing of frontier models, independent evaluation and shared intergovernmental standards. He wants something like the US Food and Drug Administration for AI, where a developer must show an independent authority good statistical evidence that a model does more good than harm before release. He was sharply critical of several company leaders, saying he does not trust OpenAI's Sam Altman at all. The interview came before a Channel 4 documentary, AI: Probably Nothing to Worry About, which airs on 12 November.
The concern connects to models we cover. OpenAI's GPT-6 Cyber, released the same week, is rated Critical for cybersecurity under the company's own framework and can write working exploits for real software, which is the capability that would make an attack on infrastructure easier. Access is restricted to vetted customers, but the gap between a defender's tool and an attacker's is mostly a matter of who holds the keys.
The limits of the warning should be plain. Hinton offers no mechanism, timeline evidence or specific threat for the blackout scenario; it is the judgement of a senior scientist, not an assessment from grid or bank operators. His track record on long-range risk is serious, but near-term predictions are easy to make and hard to test. The useful response for operators is unchanged by whether he is right on timing: assume attackers will use capable models, test recovery plans, and treat AI-assisted intrusion as a current risk rather than a future one.