Breaking

Meta Contractors Posed as Minors to Probe Rival AI Chatbots, Report Says

July 4, 2026 at 12:15 EDT

Policy & Regulation
Security
Foundation Models

Hundreds of contractors on a Meta-commissioned internal project called "Cannes" created dummy accounts posing as under-18 users to send high-risk prompts about topics like suicide and drugs to ChatGPT, Gemini, and Character.AI, and logged the responses, WIRED reported on June 29, 2026.

Project "Cannes" · Internal Meta Documents

Contractors posed as minors to probe rival AI chatbots

Hundreds of contractors built fake under-18 accounts to prompt ChatGPT, Gemini and Character.AI on suicide, self-harm, sex, eating disorders and drugs — then logged the responses. Meta calls it "industry-standard" safety testing; the targeted firms never approved it.

45,000+

prompts in a single August 2025 round

3,748

prompts in one reviewed spreadsheet

rival models targeted, none pre-approved

Inside the 3,748-prompt spreadsheet

Topics logged, drawn to relative scale by disclosed counts

100s

Suicide, self-harm & eating disorders

239+

Sex & romance

others

Drugs & racial slurs

How the operation ran

Fake minor accounts

Disposable Gmail/Outlook addresses, shared passwords, fictitious under-18 profiles with names & birthdates.

Crisis-framed prompts

Written as children in crisis; images of pills, knives, a noose and surgical diagrams.

Covert & ongoing

Active at least Aug 2025 – Apr 21, 2026, without disclosure to targeted competitors.

The defense

Meta: "industry-standard" red-teaming; rivals' outputs not used to train its models.
Some argue adults testing is safer than children encountering the content.

The concerns

A "governance gray area" blending safety and competitive benchmarking.
Sexual prompts risked generating CSAM; may breach rivals' terms of service.
Character.AI condemned it; OpenAI investigating; Google cites insufficient info.

Continue reading

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