Mark Zuckerberg has decided to vet his rivals using methods that resemble a covert operation more than industry-standard compliance. Codenamed "Cannes," the project has nothing to do with film festivals and everything to do with a massive stress test of safety systems at OpenAI, Google, and Character.AI. As revealed by WIRED, Meta hired hundreds of contractors through the firm Covalen to simulate the crisis-driven behavior of minors.
The method is as blunt as it is cynical: operators created fake accounts with birthdates under 18 and bombarded chatbots with prompts regarding suicide, narcotics, and eating disorders. The scale of the operation is striking:
In a single round of testing in August 2024, contractors sent over 45,000 queries. Every "successful" safety breach was manually logged into spreadsheets. Researchers intentionally bypassed official APIs and vulnerability reporting channels.
"AI safety is becoming a new arena for dirty games, where child protection serves as a convenient legal smokescreen for industrial espionage."
Meta defends its actions as routine benchmarking aimed at understanding industry standards. However, Google and Character.AI failed to appreciate the enthusiasm, stating that the tests were conducted without their consent and in violation of their terms of service. From our perspective, this looks like an aggressive hunt for kompromat to fuel regulatory pressure. Project Cannes appears particularly ironic given Meta's own track record; not long ago, internal guidelines reportedly allowed for rather suggestive interactions between the company's AI personas and teenagers.
Instead of a transparent audit, the industry has seen a precedent where vulnerable scenarios are weaponized. Meta positions this as a contribution to collective safety, but in practice, it is classic competitive intelligence. The tech giant essentially organized a raid on its rivals' failures, preferring to stay in the shadows until the data leaked to the press.