Microsoft's EXTRA program funds 18 university labs to find what internal red teams miss
Microsoft's EXTRA program is giving unrestricted grants to 18 university labs across six continents to hunt frontier AI failure modes that internal teams cannot reliably surface. It is an institutional admission that no lab can adversarially test its own models — and a bet that independent, funded outsiders can.
The design choice that matters is 'unrestricted.' Grants with no deliverable control and no publication veto are what you offer when you want adversaries, not contractors — researchers with the independence to report failures the funder would rather not hear. Microsoft is effectively buying the criticism it cannot generate internally, and structuring the purchase so the criticism stays credible.
Six continents is not a diversity flourish; it is methodology. Failure modes surface differently across languages, cultures, and threat environments, and a red team drawn from one country's talent pool probes one country's assumptions. Distributing the adversarial search geographically is how you find the failures that homogeneous internal teams are structurally blind to.
The program also formalizes a shift in where safety testing lives. Regulation now mandates adversarial testing for frontier systems, autonomous red-team agents are beating human operators on black-box challenges, and the labs' own evaluations keep being shown not to predict deployed behavior. External, funded, independent red-teaming is becoming the layer the industry converges on — because every alternative has publicly failed.
Tech Times — Microsoft funds 18 university labs to fix AI safety testing's blind spots → · Tech Times — AI safety evaluations are not safety certificates: formal analysis →