Story perspectives

Meta Deployed Fake Teens, Sent 45k Risky AI Prompts

7/1/2026

Summary

Meta Fake Prompts

Meta ran a covert program that used fake under-18 accounts to send more than 45,000 high-risk prompts to AI chatbots.
Contractor Covalen directed workers to target ChatGPT, Gemini, and Character.AI with prompts about suicide, sex, and drugs, and in August 2025 logged each reply in spreadsheets.
WIRED said the program was active on April 21, 2026, prompting OpenAI, Google, and Character.AI to state the testing violated their terms and was not authorized.

Summary

Meta Defends Testing

Meta publicly defended its AI safety testing, calling it routine and denying that rival AI outputs were used to train its models.
Internal records show the project, codenamed “Cannes,” remained active at least until April 21, 2026.
Prompts were submitted in languages beyond English, including a French query about bullying victim Jamey Rodemeyer.
Meta emphasized that the competitor benchmark data collected would not be incorporated into its own AI training pipelines.

1 of 2

Meta Fake Prompts

Meta ran a covert program that used fake under-18 accounts to send more than 45,000 high-risk prompts to AI chatbots.
Contractor Covalen directed workers to target ChatGPT, Gemini, and Character.AI with prompts about suicide, sex, and drugs, and in August 2025 logged each reply in spreadsheets.
WIRED said the program was active on April 21, 2026, prompting OpenAI, Google, and Character.AI to state the testing violated their terms and was not authorized.

1 / 2