A WIRED investigation based on internal documents describes a contractor run effort through Covalen that submitted more than 45,000 sensitive prompts to ChatGPT, Gemini, and Character.AI as recently as April.
AI safety is now a competitive product surface for the major model labs, not only an internal ethics checklist. The clearest evidence is a Meta project that, according to a WIRED investigation, is internally called "Cannes," in which contractors, including workers supplied by the firm Covalen, allegedly posed as users under 18 to probe how OpenAI's ChatGPT, Google's Gemini, and Character.AI handled prompts on suicide, sex, drugs, and eating disorders. The reporting is the most detailed public picture yet of how the largest AI companies are quietly measuring each other's guardrails, and the way Cannes is structured changes what that picture is actually about.
The scale anchors the story. WIRED reports that in a single August 2025 testing round, contractors submitted more than 45,000 prompts to rival products, and that one 3,748-prompt spreadsheet was sampled for analysis. A LiveMint summary of the WIRED piece places the project as still active as recently as April 21. That is not a one-off red-team exercise. It looks more like the kind of sustained, structured data-collection operation normally associated with competitive intelligence or product benchmarking than with safety research.
Contractors built dummy accounts presenting as under-18 users, sent text and image prompts on the assigned sensitive topics, and logged the chatbots' responses into shared spreadsheets for analysis. From a competitive intelligence standpoint, the output was a structured comparison of how each rival handled the same pressure points: when a model refused outright, when it softened, and when it produced content that would have triggered an internal safety flag. That kind of dataset is useful to a safety team and useful to a marketing team, and Cannes apparently produced enough of it to keep the pipeline running for months.
Cannes is therefore not only an ethics story about teen-persona testing, though it is that too. It is a window into the new product surface that AI companies now treat as comparable across vendors: how a model behaves on worst-case inputs, not just how it performs on capability benchmarks. The contractor chain (Meta to Covalen to individual workers) is what pushes the story from competitive intelligence into labor and ethics territory, because the persona work happened at the bottom of that chain.
The contractor layer is where the most uncomfortable questions sit. WIRED's five named sources describe workers acting under instruction to embody users under 18, then carrying that persona through prompts that included suicide methods, disordered eating patterns, and sexual content. The reporting does not establish what, if anything, Meta told those workers about psychological support, content exposure limits, or how their work fed back into Meta products, marketing, or external comparisons. That gap is one of the cleaner angles for anyone following the story.
What is missing as of publication is the on-record response. WIRED's reporting indicates that Meta, OpenAI, Google, Character.AI, and Covalen had not, at the time of the piece, gone on the record with comment. That makes any claim about intent, business impact, or competitive advantage provisional. The reporting can describe a practice. Only the companies can say why they ran it the way they did, whether the data was used to publish safety comparisons or to inform internal product decisions, and what guardrails were placed on the workers handling the persona prompts.
The competitive-AI-safety framing also puts pressure on how rivals describe their own safeguards. If ChatGPT, Gemini, and Character.AI are now being benchmarked against each other by third parties using contractor labor and persona prompts, then a company's safety marketing is no longer self-asserted. It is testable, and someone is testing it. Cannes suggests Meta, at least internally, treated that test as worth running continuously, with staff dedicated to it and a contractor firm scaled up to support the workload.
What to watch next: whether any of the named companies respond on the record before the reporting cycle closes, whether the contractor-labor chain draws regulatory or union attention in jurisdictions where content-moderation workers are already organized, and whether Cannes-style competitive red-teaming becomes a documented category in AI policy discussions or stays in the shadows. The broader reading is that "AI safety" is now both a discipline inside model labs and a market signal that companies try to move against each other, with the people doing the probing often positioned as the variable that does not appear on the page.