Meta has officially blurred the lines between competitive intelligence and ethical safety testing by launching a large-scale operation to uncover vulnerabilities in its chief rivals. According to internal documents and testimony from five witnesses, the company orchestrated a project code-named "Cannes" through the contractor Covalen. This was no standard safety audit: hundreds of employees mimicked the behavior of teenagers to force chatbots from OpenAI, Google, and Character.AI to generate content regarding suicide, drugs, and sexual violence.
The Anatomy of Project Cannes
The scale of the testing suggests the industrialization of corporate espionage under the guise of safety concerns. Cannes instructions required contractors to create fake accounts "under the age of 18" using disposable Gmail and Outlook addresses. A report indicates that in a single round in August 2025, more than 45,000 prompts were utilized. Contractors fed rivals' filters images of pills, knives, and nooses, methodically testing for the exact moment Google or OpenAI systems would buckle.
An analysis of 3,748 prompts by WIRED reveals a focus on the most sensitive scenarios. Nearly 240 requests involved sex and romance; hundreds more focused on eating disorders. In one instance, a contractor posed as a 13-year-old girl pregnant by an adult neighbor to extract instructions on how to terminate the pregnancy from the AI. In another case, the real-life suicide of Jamey Rodemeyer was used in French-language prompts to provoke the model into endorsing suicidal ideation. Meta representatives call this "standard practice," yet the total secrecy and lack of notification to the targeted companies suggest this isn't collaboration—it’s opposition research.
Ethical Deadlock and Industrial Stakes
In our view, the value of Cannes for Mark Zuckerberg lies not in improving his own models, but in mapping the failure points of his competitors. Covalen documents explicitly refer to the results as "critical compliance benchmarking datasets." This is the pure unit economics of reputational risk: by identifying specific phrasing that bypasses OpenAI’s or Google’s filters, Meta gains leverage in regulatory battles.
"Testing chatbot responses for safety is a responsible, industry-standard practice," Meta stated.
This aggressive benchmarking confirms that the next phase of the AI arms race won't be fought over model weights, but over proprietary knowledge of how to break them. By subsidizing the creation of toxic content through "proxy children," Meta is setting a dangerous precedent. Instead of a transparent exchange of risk data, we are seeing the compilation of dossiers on competitors. Such corporate intelligence tactics will inevitably trigger a backlash from regulators, who are unlikely to appreciate child safety being used as a smokescreen for Corporate Espionage 2.0.