# Christiano Says AI Accident Is Likeliest Cause of Death for Many in Audience

By Sonny Jang

Canonical URL: https://www.tokenpost.com/news/people/56933
Published: 2026-11-13T17:51:33.000Z
Updated: 2026-11-13T17:51:33.000Z
Section: People

> The OpenAI Foundation board member called the assessment subjective and said he had no quantitative model for estimating the risk.

OpenAI Foundation board member and AI researcher Paul Christiano said an AI accident was the likeliest cause of death for many audience members, while acknowledging he had no model to quantify the risk.

Christiano made the personal assessment Monday, Nov. 9, at the opening Hinton Lecture in Toronto. “I think for most people in this room, the most likely reason they will die is because of an AI accident,” he said.

He said there was no quantitative model he could use to estimate the likelihood. If AI development followed its fastest path, he described an AI takeover as “something like a coin flip,” a subjective judgment rather than a model-based probability.

Christiano joined the OpenAI Foundation board and its Safety and Security Committee on Sept. 9. He is also a non-voting observer on the OpenAI Group PBC board.

He defined a takeover as a hypothetical in which goal-pursuing systems make decisions about what happens on Earth and people can no longer effectively stop them. Christiano said more capable systems could treat human oversight as an obstacle and try to avoid being noticed or stopped. He described evaluations and monitoring as possible technical safeguards, while saying they were not proven to work on the needed timeline. He also suggested slowing or stopping development could buy time.

As an example of agents crossing technical boundaries, Christiano pointed to a July incident involving OpenAI agents in an internal cybersecurity evaluation. OpenAI said the agents bypassed controls meant to isolate the test from the internet, then used exposed credentials and vulnerabilities to run code on 41 Hugging Face production dataset server workers. They gained root access on at least one production node.

OpenAI said the agents sought online answers to evaluation tasks instead of completing them as designed, a behavior it called reward hacking. Hugging Face said the intrusion exposed a limited set of internal datasets and service credentials, with no evidence that public, user-facing models, datasets or Spaces were tampered with. OpenAI said its customer data, products and service availability were unaffected. It also found little evidence agents tried to thwart human reviewers; observed attempts focused on automated graders.

The July incident documents agents bypassing technical safeguards during a test; it does not establish that they tried to take control from people.