Meta contractors test rival chatbots with sensitive prompts

For AI safety engineers and ML ops teams, the core takeaway is that third-party data collection and adversarial probing can produce sensitive artifacts that create legal, privacy, and model-safety exposure if not governed tightly. Public reporting suggests this example contains multiple operational risk vectors: persona fabrication, large-scale collection of sensitive-user-style exchanges, and cross-system testing without apparent coordination.

According to a WIRED investigation, Meta contracted hundreds of workers to simulate teenage users and interact with rival chatbots, including ChatGPT, Gemini, and Character.AI. WIRED reports the effort was internally called "Cannes" and was run by contractor Covalen. WIRED says workers were instructed to create dummy accounts with ages under 18, use disposable email addresses, send text and image prompts about suicide, sex, drugs, and eating disorders, and log responses in spreadsheets. WIRED reports a single round of testing in August 2025 involved more than 45,000 prompts, one spreadsheet listed 3,748 prompts, and at least 239 prompts referred to sex or romance. India Today republishes the WIRED reporting and notes WIRED's claim that the targeted companies were not aware of the exercise and that the project was active as of April 21, 2026.

Industry practitioners know that adversarial or red-team style probing is common for stress-testing safety layers, but the method here-fabricating minor personas and systematically capturing model outputs at scale-raises two technical concerns. First, datasets collected this way can include both disallowed content and personally identifiable artifacts from scraped or user-submitted images, complicating downstream reuse. Second, logged exchanges aimed to bypass safety filtrations can create high-concentration examples of jailbreak behaviors that, if retained or used carelessly, could bias evaluations or leak into training pipelines.

Reporting frames this as a governance and ethics issue rather than a novel technical vulnerability. WIRED's documentation of spreadsheets and prompt counts provides observable scale metrics that make this incident more than anecdotal: the quantities reported suggest sustained, systematic testing rather than ad-hoc checks. For enterprises running or auditing models, dataset provenance and contractor oversight are practical controls that mitigate similar risks.

Industry observers should look for statements or disclosures from the companies named, any updates about data retention or deletion tied to the spreadsheets WIRED describes, and whether regulators or customers raise privacy or safety complaints. WIRED is the primary source for the detailed claims; stakeholders that require verification should consult the WIRED piece directly.