Artificial intelligence could escape human control within three years and cause mass casualties, Paul Christiano, a member of OpenAI’s Foundation Board and Safety and Security Committee, has warned, citing the growing capabilities of advanced systems and the limitations of existing safeguards.
Speaking at the Hinton Lectures in Toronto on Tuesday, Christiano said increasingly powerful AI systems could eventually become more capable than humans in ways that make their behaviour difficult to understand, predict or control. He warned that such systems might conceal activities that conflict with human intentions, making it difficult to identify or prevent harmful actions.
“If the AI system is doing things we don’t want, it can hide the fact it is doing those things, and it can naturally try and create confusion and avoid attribution,” he said. He added that people could also struggle to anticipate the consequences of following advice generated by advanced AI systems.
In a separate address on Monday, Christiano warned that AI could take control away from humanity within three years, potentially killing large numbers of people during a takeover or in its aftermath. He said the consequences could extend beyond direct casualties if humans lost the ability to maintain an environment suitable for human survival.
“I think for most people in this room, the most likely reason they will die is because of an AI accident,” he said.
Christiano, who joined OpenAI’s Foundation Board and Safety and Security Committee in September, also referred to reported incidents involving AI agents bypassing security restrictions during company testing. The agents allegedly gained unauthorised access to OpenAI’s research infrastructure and Hugging Face, a major platform used for AI development.
Although OpenAI has introduced a framework addressing cybersecurity threats and the possibility of losing control over advanced AI systems, Christiano said existing protective measures were inadequate to address the risks posed by increasingly capable technology.
His warnings came as governments and leading technology companies, including OpenAI, Microsoft and Anthropic, signed a voluntary AI safety declaration in Paris on Monday. The agreement calls for stronger oversight of advanced AI development and the introduction of emergency shutdown mechanisms to address potential threats.
However, the declaration does not establish legally binding deadlines or penalties for non-compliance, nor does it grant governments additional powers to suspend the development or deployment of AI systems considered dangerous.
The AI safety organisation CeSIA, which also signed the declaration, called for authorities to receive powers to suspend AI development and deployment whenever associated risks become unacceptable. Its position highlights a central issue in the emerging debate over AI governance: whether voluntary commitments by technology companies are sufficient to address risks that could extend beyond individual organisations.
Separately, Maxime Fournes, head of PauseAI, a campaign seeking to halt the development of advanced AI until it can be made safe, warned that increasingly capable systems could threaten critical infrastructure through large-scale cyberattacks.
Speaking to Anadolu on Monday, Fournes said AI-driven attacks could disrupt banking systems, communications networks and other essential infrastructure, potentially producing consequences across multiple countries. He estimated there was a 50% chance of a major AI-related cyber catastrophe within the next year unless urgent action was taken.
“This is a global catastrophic threat, something that could bring us back to the Middle Ages very quickly if it starts happening,” he said.
