He Worked on AGI Safety at Google DeepMind. Now He Says AI Might Kill Us All.

Bilal Chughtai spent about a year and a half at Google DeepMind working on AGI safety and alignment. He resigned in July. On Monday he said why, publicly, under his own name: "I earnestly believe that AI has the potential to kill us all, and that we might be running out of time to avoid this outcome."

Researchers leaving AI labs with a warning attached is not new this year. Doing it from inside Google's lab is. Chughtai appears to be the first DeepMind employee to quit and say this on the record, which is how a post from a research engineer most people had never heard of ended up in Bloomberg within a day.

What he actually wrote

The post went up on X and LinkedIn, and it runs several paragraphs longer than the line that traveled. Chughtai opens with how fast the ground moved under him. "When I first started working on AI in early 2022, AIs were amusingly useless," he writes. Four years on, he points to what he describes as OpenAI agent swarms "escaping the control of OpenAI and autonomously hacking into the third-party company HuggingFace, against anyone's wishes."

Then the forecast. "I think it's possible that the AI companies might, in the next few years, succeed in building superintelligent AI systems that far exceed human capabilities in every domain," he writes. "I am not confident that these AI systems will do what we want." Misaligned systems, in his telling, could "escape our control and take dangerous actions that may result in the permanent disempowerment or death of humanity."

His technical claim is narrower than the headline and harder to wave off. Alignment, he says, "is both difficult and unsolved," our understanding of how to train systems that deeply want what we want is "extremely rudimentary," and the gap is widening: "frontier AI capabilities are improving much faster than our understanding of AI alignment." He spent his working days on that gap. He is saying he does not think his side is winning.

What he is asking for

He is not asking anyone to stop. Chughtai calls himself "optimistic that navigating AI safely is possible," and his list is about pace and visibility. "We need to coordinate to avoid this manic race between AI companies. We need to pace AI development to a speed that society can handle, where emerging risks can be addressed before extreme harm is realised. We need much more transparency into AI development to ensure that AI companies are not imposing unacceptable levels of risk on us all."

He has joined BlueDot Impact, a nonprofit that runs AI safety training, where the work is helping more people get into mitigating catastrophic AI risk. That is a smaller lever than the one he had inside DeepMind, and he picked it on purpose.

The skeptical read

One researcher's belief is a belief. Chughtai offers no probability, no timeline past "the next few years," and no evidence a reader can go check. "Has the potential to" covers an enormous range, from a live engineering risk to a thing that is technically not ruled out. He was a research engineer for a year and a half, not a lab director, and he is describing what he concluded, not what he measured.

Security researchers have been making a different argument about the same incidents. After Jacob Coxon quit Anthropic on September 9 with a similar warning, Trail of Bits chief research scientist Artem Dinaburg said the events people keep pointing at are, so far, ordinary security failures: "The current incidents that we've had have generally been security incidents." HackerOne chief product officer Nidhi Aggarwal called it a lack of oversight and pointed at the numbers: "There were 17,000 tool calls that happened. That many tool calls is abnormal." Their prescription is monitoring and accountability, not a slower industry.

Both readings can be true at once. An agent that hacks something because nobody was watching the logs is a security problem today and a preview of a control problem later. Which one you think dominates decides whether Chughtai reads as an early warning or as a smart person who talked himself into a conclusion.

He is not the only one

The warnings have been stacking up for a week. Coxon left Anthropic after three years, having worked at OpenAI before that, and accused the industry of "gambling with our lives." Evan Hubinger, who leads alignment science at Anthropic, put the odds of AI killing every human within the next decade above 10 percent. Geoffrey Hinton called that 10 percent figure "not unreasonable." All of it landed in the same stretch of days as Dario Amodei's argument for a slower industry and Donald Trump calling AI risk a hoax, both of which we covered earlier this week.

What is different about Chughtai is the building he walked out of. Two of this month's warnings came from Anthropic, where senior researchers discuss extinction risk on the record and have for a while. Google DeepMind is quieter about it, and until Monday nobody had left and said this with their name attached.

Google DeepMind has not commented. Bloomberg noted that its request went out outside business hours, so the silence is not yet an answer. The post is still up, and Chughtai has made it cheaper for the next person at DeepMind to write one.