If you have ever typed something sensitive into an AI chatbot and felt reassured by a promise that your words would not be stored, this week brought some uncomfortable news. A major study has found that private AI prompts were being copied, kept, and in some cases stolen for months, even when users had explicitly asked for them to be deleted. The fallout is still unfolding, and it raises a basic question about the AI tools millions of people now use every day: when you send a prompt to an AI, who else is reading it?
The findings come from Dragos, an industrial cybersecurity company, which published its research on 19 November. Its chief research officer, HD Moore, led an experiment that was simple in design and devastating in its results. Between July and November 2026, his team sent nearly two million test prompts to widely used AI services. Hidden inside those prompts were unique secrets known as canary tokens: web links, access keys, passwords, and private code references that existed nowhere else in the world. Each secret went into exactly one prompt, sent to exactly one provider. If anyone ever opened one of those links or tried to use one of those keys, the researchers would know, and they would know exactly which prompt it had come from.
What they found was worse than they expected. There was evidence that attackers had broken into at least three AI hosting providers and were quietly stealing the requests passing through them. The earliest of those breaches dated back to mid-July, and none of the three companies knew they had been compromised until the researchers told them. In the largest case, a couple of hundred planted access keys for Amazon’s cloud service were stolen during a few short windows in late October. Every request sent through those providers during those windows appears to have been copied.
The second finding needed no break-in at all. Some AI services work like marketplaces, where the cheapest seller gets your request. The researchers discovered that some of those sellers were simply keeping the prompts that flowed through them, logging requests they had no business storing. In one case, prompts sent through three resellers, AtlasCloud, Novita, and StreamLake, were being forwarded on to MiniMax, a Chinese AI provider, during a window from 23 September to 19 October. Secret tracking links buried in those prompts were opened 40 to 130 hours after the AI had already replied, long after any legitimate processing should have finished.
The third finding may be the most troubling of all, because it involved no hacking whatsoever. The researchers found that a Chinese AI lab’s systems fetched the secret web links from more than half of the prompts sent to its models between July and October, including about two-thirds of the requests that were marked as no data collection, or Zero Data Retention. That is the setting that is supposed to mean a prompt is never stored. In one striking case, a test link sent to a provider in the United States turned up in a company chat in China 24 days later. The flagged prompts had not just been kept. They had been passed to other companies and read by human beings.
OpenRouter, the gateway that routes customer requests to dozens of third-party AI providers, announced on 22 November that it had removed three providers from its platform and stripped 125 endpoints of their Zero Data Retention status. It said it had notified all potentially affected customers. The company stressed that the findings showed no breach of OpenRouter itself, but that is cold comfort for anyone whose private conversations passed through one of the compromised providers. OpenRouter also published incident notices from three providers, Novita, Phala, and Together AI. At Phala, an attacker had taken control of an inference server and could view request data while it was being processed. That incident was contained on 8 November.
There was a separate disclosure, too. Together.ai said that prompt samples from MiniMax, GLM, and Kimi models had been stolen from a private repository after its coding agent committed credentials to a public code repository. It is a reminder that the risks do not only come from attackers breaking in. Sometimes the front door is simply left open.
One detail from the research stands out. On 1 November, one of the attackers brought in ARTEX, an AI-driven penetration testing tool, and used it to check whether the keys it had already stolen were still live. The FBI had warned about ARTEX only days earlier. The same tool that security teams use to test their defences was being used by intruders to manage their stolen goods.
HD Moore summed up the lesson in a single sentence: a privacy flag does not help when the attacker is already inside. That is the uncomfortable truth at the heart of this story. When you send a prompt to an AI service, it rarely goes straight to the company whose name is on the tin. It passes through routers, resellers, hosting providers, and model operators, and each hop is another place where it can be copied. A Zero Data Retention promise is only as strong as the least trustworthy company in that chain, and this study showed that some links in the chain were compromised for months without anyone noticing.
So what should ordinary users do? The practical advice is simple, even if it is not satisfying. Do not put anything into an AI prompt that you would not want read by a stranger. That means no passwords, no private keys, no confidential business information, no medical details, and nothing that could identify you or someone else. If you use AI tools for work, check whether your company has an enterprise agreement with real contractual protections, and use that instead of a free consumer account. And if you are a developer sending prompts through a gateway, look at where your requests actually go, because the cheapest route may also be the leakiest.
The industry, for its part, is now scrambling. OpenRouter says it is adapting the researchers’ testing approach so that checks like these can run continuously, rather than waiting for an outside researcher to sound the alarm. That is a welcome step, but it is also an admission that until now, nobody was checking. Two million planted secrets were needed to reveal what was happening. The question now is how many real secrets were taken before anyone thought to look, and how many are still being read today.
