A group of mathematicians has launched the Mathematical AI Safety Institute (MAISI), a research organisation that aims to develop rigorous mathematical proofs of AI safety, in the way cryptographers prove that an encryption scheme cannot be broken. Its scientific director is Jacob Tsimerman, a 2026 Fields Medalist and mathematics professor at the University of Toronto, who is also joining OpenAI's safety team, according to The Decoder. The executive director is Andrew Critch, who holds a PhD in mathematics from UC Berkeley, and the chief operating officer is Jake Ross, previously COO of Encultured AI.
A mathematical approach to safety
MAISI describes itself on its website as an independent non-profit developing mathematical foundations for the safety of powerful AI systems, and states that "humanity deserves a mathematically principled account of existential safety for new AI technologies." Its scientific advisory panel includes Ravi Vakil, the Robert Grimmett Professor at Stanford and president of the American Mathematical Society since 2025; Timothy Gowers, a 1998 Fields Medalist who holds the Chaire de Combinatoire at the Collège de France; Geoffrey Irving, co-founder and chief scientist of Resolution; and Paul Christiano, director of the Alignment Research Center. Its board includes the University of Toronto mathematicians Arul Shankar and Yevgeny Liokumovich.
The institute's stated goals include developing tools such as zero-knowledge proofs, which could let AI systems demonstrate their integrity without exposing an AI lab's trade secrets, establishing clear theoretical definitions of what counts as safe, and finding ways to prevent unwanted outcomes when multiple AI agents interact rather than operating in isolation. Tsimerman told the New York Times, in reporting relayed by The Decoder, that the field needs "a much, much higher level of safety standard than we're currently getting."
Hiring and timeline
MAISI, based in the San Francisco Bay Area, is recruiting ten to thirty faculty for a first semester beginning in January 2027, and between thirty and one hundred for a special-year programme starting in September 2027. The institute does not list funders or a budget on its site.
The launch adds a distinct voice to a safety debate largely driven so far by AI labs' internal safety teams and policy researchers rather than pure mathematicians. The Decoder notes a core difficulty with the approach: unlike an encryption scheme, where the property to be proved is precisely defined, AI safety problems tend to surface only in practice, and there is no agreed theoretical definition of what safe means. By framing safety as a problem for the mathematics community, MAISI is betting that formal proof techniques can eventually offer guarantees for systems whose behaviour is far harder to specify in advance.