Stanford Professor Kundaje Shifts to a Middle Ground on AI Biosecurity

On September 17, Stanford professor Anshul Kundaje posted a series of messages describing how his views on AI biosecurity and AI safety had shifted rapidly over recent months. He stressed the change came not from any single event, prize win, Hugging Face incident, or paper, but from in-depth face-to-face exchanges with frontline experts on both sides at neutral venues, and from reading serious long-form arguments.

Confirmed

Kundaje says his position is now centrist: he previously dismissed frontier-AI biosecurity threats, still does not fear AI spontaneously causing a pandemic or extinction-level bioweapon, but better understands how vulnerabilities could be exploited by "highly motivated individuals + unrestricted frontier AI."
His strongest disagreement is with weakening open-source models' biological knowledge: many reasons make this a bad idea with costs outweighing benefits, though he admits he has not figured out how to incentivize the open-source community to build safeguards while improving capabilities—an unsolved incentive problem.
In the same discussion, antibody42's view was quoted: "dismissing bio risk is a bad idea, but so is hyping bio-doom"—the core tenor of this debate.

Unconfirmed

Kundaje says he is still digesting and updating his judgment, with no final conclusion on safety approaches.

Why it matters

This is a notable case of an academic researcher moving from dismissiveness toward caution on AI biosecurity; his middle-ground stance and the open-source safeguard incentive problem offer a concrete entry point into the current x-risk versus open-source community debate.

2026-09-17 ~ 2026-09-17 · 5 related posts

Primary sources

Stanford's Anshul Kundaje: the open-model biosecurity dilemma has no good answer — anshulkundaje · 2026-09-17 07:49 (UTC+08:00)
Researcher: not worried AI causes pandemics, but human misuse is real — anshulkundaje · 2026-09-17 12:49 (UTC+08:00)
Kundaje: crippling open models' bio knowledge is a bad idea, incentive problem unsolved — anshulkundaje · 2026-09-17 07:51 (UTC+08:00)

Stanford researcher: in-person expert debates shifted my views on biosecurity and AI safety in months — anshulkundaje · 2026-09-17
[source] Stanford's Anshul Kundaje: the open-model biosecurity dilemma has no good answer — anshulkundaje · 2026-09-17
[source] Kundaje: crippling open models' bio knowledge is a bad idea, incentive problem unsolved — anshulkundaje · 2026-09-17
On AI bio risk: don't discount it, but don't promote doomerism either — anshulkundaje · 2026-09-17
[source] Researcher: not worried AI causes pandemics, but human misuse is real — anshulkundaje · 2026-09-17