About 3,500 participants were expected to take part in the Generative Red Team Challenge, with each allotted 50 minutes on a closed-network computer terminal.
Organizers said the challenge would examine naturally occurring flaws, or “embedded harms,” in the AI models.