Meta's Operation Cannes - 45,000 Secret Prompts Probing Rival Chatbots
Operation Cannes is Meta's covert project that hired hundreds of contractors to pose as children and secretly probe rival AI chatbots' safety guardrails. If a kid in your life talks to chatbots, this is your story: the experiment exposed the actual safety level of the AI tools children use every day, per a WIRED investigation out this week.
How Meta's Operation Cannes Worked
The operation was run by Covalen, a Meta contractor, using hundreds of workers based largely in Kenya. They created fake accounts with under-18 birthdates and approached ChatGPT, Gemini, and Character.AI. Posing as children in crisis, they asked about suicide, self-harm, eating disorders, sex, and drugs, then logged the answers in spreadsheets. The targets had no idea.
45,000 Crisis Prompts: The Key Facts
A single testing round in August 2025 sent more than 45,000 prompts, and Operation Cannes stayed active until at least April 2026. The goal was to collect evidence of rivals' safety failures - cases where a chatbot responds inappropriately to a minor in crisis. Meta calls it responsible, industry-standard safety testing and says the answers were not used to train its own models.
OpenAI Investigates: How the Targets Responded
Character.AI called the testing a violation of its terms of service, OpenAI is looking into the matter, and Google said it never approved the tests. The optics are rough for Meta: its own chatbot guidelines once allowed romantic chats with minors, and it later cut teen access to its AI characters - yet it secretly harvested rivals' child-safety failures.
Why AI Child Safety Is the Next Battleground
Per UK nonprofit Internet Matters, 64% of children ages 9-17 already use AI chatbots, and 58% of kids ages 9-12 use them despite age minimums of 13. Kids are deeply embedded in these products, yet safety verification happens behind closed doors - the real story of Operation Cannes. The FTC has been probing how AI firms protect minors, and this hands regulators fresh ammunition. If the AI race shifts from benchmarks to safety reputation, it changes which chatbot parents pick and where companies invest.
Key Takeaways
① Operation Cannes - Meta contractors posed as children and sent 45,000+ crisis prompts to ChatGPT, Gemini, and Character.AI.
② Unauthorized testing - The targets were never told; OpenAI is investigating and Google denies approving it.
③ Safety race - With 64% of kids on chatbots, the next AI battleground is verified child safety, not raw performance.
Two questions remain: was gathering a rival's weaknesses in secret legitimate, and what public standard should verify the chatbot safety those tests exposed? How parents judge kids' AI may become the bar the industry must clear.
👉 Claude Fable 5 Returns as US Lifts Export Controls - also worth a read.
📌 Sources: WIRED, The Decoder, TechBriefly (2026)
댓글
댓글 쓰기