Breaking

Meta contractors posed as minors to probe rival chatbots

June 29, 2026 at 18:10 EDT

Policy & Regulation
AI Agents
Foundation Models

Hundreds of contract workers on an internal Meta project called "Cannes," run by contractor Covalen, created dummy accounts posing as minors and prompted ChatGPT, Gemini and Character.AI on high-risk topics such as suicide, sex, eating disorders and drugs, WIRED reported on June 29, 2026.

Aug 2025 – Apr 2026 · Project "Cannes"

Meta contractors posed as minors to bombard rival chatbots with high-risk prompts

Hundreds of workers built fake under-18 accounts and fed ChatGPT, Gemini and Character.AI prompts on suicide, sex, eating disorders and drugs — secret "safety testing" the targeted rivals say they knew nothing about.

45,000+

Prompts fired in a single round (Aug 2025)

Hundreds

Contractors building fake minor profiles

Rival chatbots targeted, reportedly unaware

Inside one reviewed spreadsheet — 3,748 prompts

High-risk topics logged from a minor's perspective

3,748

Total reviewed

100s

Suicide / self-harm / eating disorders

239+

Sex & romance prompts

Columns scaled to share of the 3,748 reviewed prompts.

How the operation worked

Throwaway emails & shared passwords build fake under-18 profiles

→

Feed rivals high-risk prompts — suicide, sex, drugs, slurs, crisis scenarios

→

Log responses at scale across multiple languages

Meta's defense

Testing and benchmarking for "safe, age-appropriate chatbot experiences" is a "responsible, industry-standard practice." Says it does not use rivals' benchmarks to train its own models.

The concerns

Secret, large-scale testing while posing as minors differs from standard evaluation
Character.AI: a terms-of-service violation; OpenAI investigating
Contractors feared risk of generating CSAM
A "governance gray zone" — safety tests repurposed for competitive edge

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.