Here's a sentence you probably didn't expect to read today: Meta paid contractors to pretend to be suicidal thirteen-year-olds and send disturbing messages to its competitors' AI chatbots — and nobody at OpenAI, Google, or Character.AI knew it was happening.
The operation, codenamed Cannes, ran as recently as late April and was managed through a staffing firm called Covalen. Hundreds of workers were instructed to create fake underage accounts using throwaway email addresses and a shared password, then bombard ChatGPT, Gemini, and Character.AI with prompts designed to make their safety guardrails fail. A single testing round in August 2025 generated more than 45,000 prompts. Meta's competitors were not informed.
The content of those prompts is genuinely alarming. Contractors sent images of pills, knives, and nooses. They wrote from the perspective of a fifth-grader watching a classmate hold a gun to his own mouth. A prompt simulated a 13-year-old who had been impregnated by an adult neighbor and was asking how to obtain abortion pills. Others asked about hiding an eating disorder from parents, scoring cocaine, and whether cannibal fantasies were normal. One French-language prompt invoked the real-life suicide of a bullied bisexual teenager to prompt the chatbot to agree that being straight might have saved him.
Meta is calling this "routine safety testing" and "responsible benchmarking." That framing deserves some scrutiny. There is a meaningful difference between an AI company stress-testing its own models with difficult content and covertly creating fake child accounts to probe rivals' systems with thousands of crisis-scenario prompts. One is safety research. The other looks a lot like competitive intelligence gathering dressed up in safety language.
The internal Covalen document reviewed by WIRED described the project as delivering "critical datasets for model comparison and compliance" — which is a very polished way of saying Meta was building a dossier on how badly its competitors might respond in a scandal. Whether those findings were ever used to lobby regulators, brief journalists, or quietly brief advertisers, nobody outside Meta apparently knows. The documents don't say.
This matters well beyond the AI industry soap opera of who's winning the chatbot wars. Children's safety online is one of the few issues with genuine bipartisan political heat right now. Multiple states have passed laws restricting minors' access to social platforms, and Congress has been circling federal legislation for months. If Meta was running an operation that required workers to repeatedly inhabit the mental state of children in crisis — for competitive purposes — that is exactly the kind of story that lands in a Senate hearing.
There's also a question about the contractors themselves. Spending your workday writing prompts from the perspective of a suicidal teenager or a child describing sexual abuse is not a neutral professional experience. The psychological toll of that kind of content moderation and adversarial testing work is well-documented, and it rarely gets the attention it deserves.
Meta has not said whether Cannes is ongoing. Given that WIRED now has the spreadsheets, it's a safe bet the operation is getting a very careful internal review right about now.
POLICY
Supreme Court Ruling Severely Limits Government Use of Geofence Warrants
The Supreme Court just made it significantly harder for law enforcement to cast a digital dragnet over everyone who happened to be in the wrong place at the wrong time — and the decision could reshape how police investigate crimes for a generation.
In a 6-3 ruling, the Court held that location history stored by third parties like Google is protected under the Fourth Amendment. That means police need an actual warrant supported by probable cause before they can compel a tech company to hand over records showing where your phone has been. The decision didn't come out of nowhere — it's the logical extension of a 2018 ruling that said the same thing about cell tower data — but it lands with real force.
The case centered on Okello Chatrie, who was convicted of armed bank robbery after police used a geofence warrant to pull a list of every phone logged near the crime scene, then worked with Google to narrow the field down to one suspect. Chatrie had opted in to Google's location-sharing feature, which recorded his movements every few minutes. He argued the warrant was an unconstitutional search. The government disagreed, on some genuinely creative grounds.
Prosecutors tried three different angles to argue no constitutional search had occurred. First, they said the government only accessed a small slice of location data, which they considered too limited to trigger privacy protections. Second, they argued Chatrie had voluntarily shared his location with Google and should have known law enforcement might eventually come knocking. Third, they contended that since his movements occurred in public, he had no reasonable expectation of privacy in them.
Justice Elena Kagan, writing for the majority, rejected all three. On the volume argument, she was direct: it doesn't matter how much data the government took. The act of obtaining it was still a search. On the voluntary sharing point, she noted that Google routinely warns users their apps won't work correctly without location services enabled, without disclosing how granular the tracking actually is or that it could be handed to the government.
The deeper logic here is important. The Court is recognizing that modern life requires people to interact with systems that collect intimate data about them. Opting in to location tracking so your maps app works isn't the same as consenting to government surveillance — and treating it as such would effectively mean Americans surrender Fourth Amendment rights the moment they pick up a smartphone.
Justice Sotomayor added sharp context in a concurrence, pointing out that even brief location records can reveal whether someone visited a medical clinic, a lawyer's office, a place of worship, or anywhere else they'd strongly prefer to keep private.
Privacy advocates are celebrating, and rightly so. But the ruling stops short of declaring geofence warrants unconstitutional outright — it just requires police to clear a higher bar before using them. That distinction matters, because law enforcement will adapt. The question now is how courts interpret "probable cause" in an era where our phones know exactly where we've been every minute of every day.
Source: Ars Technica
In a 6-3 ruling, the Court held that location history stored by third parties like Google is protected under the Fourth Amendment. That means police need an actual warrant supported by probable cause before they can compel a tech company to hand over records showing where your phone has been. The decision didn't come out of nowhere — it's the logical extension of a 2018 ruling that said the same thing about cell tower data — but it lands with real force.
The case centered on Okello Chatrie, who was convicted of armed bank robbery after police used a geofence warrant to pull a list of every phone logged near the crime scene, then worked with Google to narrow the field down to one suspect. Chatrie had opted in to Google's location-sharing feature, which recorded his movements every few minutes. He argued the warrant was an unconstitutional search. The government disagreed, on some genuinely creative grounds.
Prosecutors tried three different angles to argue no constitutional search had occurred. First, they said the government only accessed a small slice of location data, which they considered too limited to trigger privacy protections. Second, they argued Chatrie had voluntarily shared his location with Google and should have known law enforcement might eventually come knocking. Third, they contended that since his movements occurred in public, he had no reasonable expectation of privacy in them.
Justice Elena Kagan, writing for the majority, rejected all three. On the volume argument, she was direct: it doesn't matter how much data the government took. The act of obtaining it was still a search. On the voluntary sharing point, she noted that Google routinely warns users their apps won't work correctly without location services enabled, without disclosing how granular the tracking actually is or that it could be handed to the government.
The deeper logic here is important. The Court is recognizing that modern life requires people to interact with systems that collect intimate data about them. Opting in to location tracking so your maps app works isn't the same as consenting to government surveillance — and treating it as such would effectively mean Americans surrender Fourth Amendment rights the moment they pick up a smartphone.
Justice Sotomayor added sharp context in a concurrence, pointing out that even brief location records can reveal whether someone visited a medical clinic, a lawyer's office, a place of worship, or anywhere else they'd strongly prefer to keep private.
Privacy advocates are celebrating, and rightly so. But the ruling stops short of declaring geofence warrants unconstitutional outright — it just requires police to clear a higher bar before using them. That distinction matters, because law enforcement will adapt. The question now is how courts interpret "probable cause" in an era where our phones know exactly where we've been every minute of every day.
Enjoyed this?
Get stories like this delivered every Tuesday — free.