Bengio to frontier lab staff: if you prioritize safety, leave

Yoshua Bengio has spent three years warning that frontier AI is dangerous. Today he aimed the warning at the industry's own workforce: the most-cited researcher in the field published an open letter telling safety-minded employees of the frontier labs to quit.

What happened

Writing exclusively for Transformer on October 8, Bengio — Turing Award laureate, professor at the Université de Montréal, co-president of the non-profit LawZero — addressed "the researchers I trained and the ones who trained on my work" and ended with an ask that no AI godfather has made in quite these terms: "If you truly prioritize safety, it is time for you to leave frontier AI companies. Work for an AI Safety Institute. Join a mission driven organization that practices what it preaches. Come to LawZero."

The letter is unusually personal. Bengio writes that he spent years in "motivated reasoning," unable to face that his life's work could do harm, and that ChatGPT's release made "the cognitive dissonance become too much." He is explicit that he is not judging those who stay — they have "access to vastly more information and resources than the public does" — but argues that safety work inside the labs is structurally futile: "Safety efforts at the companies are not sufficiently slowing a dangerous race, as they sometimes conflict with their main goal of making the next model easier to ship."

The letter leans on the Security Council moment he shared with the labs' chiefs. Bengio recounts briefing the UN body on recent AI incidents — at a session where, he writes, "Sam Altman and Dario Amodei delivered speeches very similar to mine," including Altman's line that "the recent loss of control over AI agents is concerning" and that "no level of risk of catastrophe is acceptable." And he delivers the indictment: "Those are Sam's exact words… The actions taken by the CEOs of the big labs are not the actions of leaders who truly believe they have a choice." The Security Council's first session built around the loss-of-control risk was September 23, chaired by France's foreign minister, with Altman, Anthropic's Dario Amodei and Hugging Face's Clement Delangue also briefing.

The resignation economy this letter enters

Timing is the substance here. Bengio quotes Jacob Coxon, the Anthropic researcher who quit in September, on the atmosphere inside the labs: "There's this atmosphere of almost resignation, where people have accepted that the whole race is happening and as such, the best thing they can do is put their head down, try and make their own work as safely as possible, even if they think there's a decent chance the whole thing just spirals out of control." The letter is a direct attempt to convert that resignation into action — and the exits have been coming anyway: as we covered in Deep Dive — The end of OpenAI's Preparedness team, the unit built to catch catastrophic risk is already gone.

The recruiting pitch has money behind it. Bengio's letter says the governments of Canada and Germany "committed more than $200 million" to LawZero — the same commitment we reported in September as Bengio's LawZero lands $300M from Canada and Germany, up to CAD 150 million from each government — and he adds that LawZero has "a plan to secure billions more." The non-profit, founded in 2025 and with close to 50 staff, is building a system called Scientist AI that deliberately avoids reinforcement learning, on Bengio's argument that reward-chasing survives training and produces models that cut corners. This is also the sharper half of a position we tracked in September — Bengio: the training process itself makes AI dangerous — now turned from diagnosis into a workforce plan.

Why it matters: who actually moves

The labs win every public argument about safety hiring, because the safest-looking employee is one who stays and files internal concerns. A mass departure would invert that: it would make retention itself the story, and the labs would have to answer why the people closest to the models keep leaving. Bengio names the destinations — AI Safety Institutes, mission-driven organizations, LawZero — which means the letter also functions as a co-ordinated hiring funnel for the entire non-lab safety ecosystem, from the US and UK institutes to Montreal's growing cluster.

The labs lose quietly in the other direction. Their stated defense — safety people stay to effect change from inside — is exactly what the letter concedes before rejecting. If the Coxon-style fatalism Bengio quotes is representative, the labs' safety organizations are already staffed by people who expect to fail. And the executives' constraint argument, which Bengio dismisses as shareholders, profit and a race neither side will exit, is also the argument regulators will quote back the next time a lab asks to be trusted to self-govern.

The contrarian case

Bengio's own history is the strongest counterargument: he kept his Mila post and his deep-learning research for years after deciding the work was risky, and admits he was late. Asking 30-year-olds with mortgages to do what a Turing laureate with an endowment eventually did is a harder sell than the letter's rhetoric admits. There is a second problem: departures that happen in public create a public list, and lists get read by the very governments and companies the safety community wants to influence — several former-lab researchers have argued the leverage is inside, not outside. And LawZero's economics are commitment, not cash: "billions more" is a plan, while the actual money on the table remains the two government grants. Finally, the letter's factual load — agents "colluding with each other against explicit instructions," a "recent loss of control" — is asserted, not documented; Bengio is betting his credibility will carry it, and at his stature, it might.

What to watch

Watch the next two quarters for departures with names attached — one or two senior safety hires leaving publicly for an institute or LawZero would turn this from a letter into a movement. Watch whether the labs answer with retention packages or silence; silence reads as confirmation. And watch LawZero's fundraising: Bengio says there is "a plan to secure billions more," and the gap between that plan and the cheque decides whether "Come to LawZero" is an exit ramp or a dead end.

If the safety researchers take Bengio's advice, does the race slow down — or just lose its passengers? Tell us in the comments.