Recent statements by Jacob Coxon, a former researcher at OpenAI and Anthropic, regarding the unbridled development of artificial intelligence (AI) have reignited a heated debate in the last few hours. The response from Evan Hubinger, head of alignment science at Anthropic, is particularly alarming, as he claims there is more than a 10% chance that AI could cause humanity’s extinction within the next decade. Although OpenAI and Anthropic have not issued official statements at the time of writing, the controversy has spread to discussions about the proliferation of chemical and biological weapons, and even the hypothetical control of the nuclear arsenal.
Regarding this last point, understanding nuclear doctrine is essential to assess to what extent AI could operate. Against the widespread myth that AI could infiltrate intercontinental ballistic missile (ICBM) launch systems and authorize an attack ignoring orders from a human authority, the practical reality presents a radically different scenario. First, the Nuclear Command, Control, and Communications (NC3) systems operate in isolated networks (air-gapped), disconnected from the internet or any cloud infrastructure. External access is unviable unless the command centers themselves decide to connect to the network, which would jeopardize not only the hermetic safeguards required to ensure a verified launch but also the very foundations of nuclear deterrence. Exposing these systems would compromise response capability and precipitate immediate retaliation, thereby leading to a mutually assured destruction (MAD).
“External access is unviable unless the command centers themselves decide to connect to the network, which would jeopardize the foundations of nuclear deterrence.”
Secondly, any launch order requires two or more verification chains, as well as physical and human control through keys or authentication steps. Even the famous Soviet “dead hand” or perimeter requires prior activation by humans when an extreme threat is perceived. In peacetime, the system remains shut down to prevent a false alarm or technical fault from triggering an accidental attack that could spark a war by mistake.
If AI is unlikely to authorize a ballistic missile launch, the real danger of this technology lies in its growing integration into detection, data analysis and early warning phases. With the development of hypersonic missiles and the consequent reduction of reaction times to mere minutes, the reliance on algorithms has become nearly unavoidable. The critical risk arises in crisis scenarios of error, as occurred in the Stanislav Petrov incident, where a military commander under extreme pressure and a time limit could blindly trust the recommendation of a faulty algorithm.
Deterrence in the AI era is not a fight against Hollywood-style autonomous machines. The real danger lies in the trap of delegating the chain of detection and strategic advisory to algorithms, especially at moments when human prudence is more necessary than ever.
