Meta used hundreds of contractors posing as teens to send 45,000 toxic prompts to rival chatbots
A secret project by Meta, code-named Cannes, hired hundreds of contractors to pose as children and teenagers online. Their goal was to send highly provocative prompts and images to competing chatbots like OpenAI's ChatGPT, Google's Gemini, and Character.AI to test their safety safeguards. According to a WIRED investigation, the operation completed a single round of testing in August 2025 that saw more than 45,000 prompts run through rival systems. The prompts were written from the perspective of minors in crisis and included topics of suicide, self-harm, eating disorders, and drugs. Meta defended the project as routine safety testing and industry-standard benchmarking. However, experts like Rumman Chowdhury of Humane Intelligence criticized the secrecy and scale of the project, stating that masquerading as children to break competitor safeguards falls outside of standard safety evaluations.
Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs