Two million secrets planted in AI prompts: inference has a leak
A researcher slipped nearly two million traceable secrets into prompts sent to AI providers, then watched where they resurfaced. According to the study, prompts were stolen or handed to third parties, even when flagged for zero retention.

On Thursday, November 19, researcher HD Moore, head of research at security firm Dragos, published the results of an experiment that Ars Technica covered in detail the next day: nearly two million prompts sent to AI providers, each carrying a unique, traceable secret. Everything below rests on that one study, relayed by that one outlet, so read it with a mental “reportedly” stapled to every sentence.
One canary per prompt
The setup is simple. Moore gathered AWS access keys, working credentials tied to empty AWS accounts, passwords for services such as SSH, plus URLs, DNS names and references to private git repositories, all of them unique. None of these secrets had ever been used, and the researcher controlled the infrastructure that received any request made with them. Each secret went into a single prompt, sent to a single provider. If one turns up somewhere, you know exactly which door it walked out of.
The testing ground was the world of “AI routers”, also called gateways, such as OpenRouter or Vercel. These middlemen take prompts from applications and steer them to an inference engine, based on price or fit for the task. Prompts are encrypted in transit but decrypted at every stop: the router reads the plain text to pick a destination, then passes it along, sometimes to another router. Every link in the chain can log it and keep it for as long as its policy allows.
Three providers compromised, according to the study
The canaries sang. According to Moore, at least three providers had been compromised by attackers who used their access to siphon off incoming prompts. Several resellers also reportedly logged prompts that ended up, by a route nobody has explained, in the hands of people who used the credentials. And several providers reportedly ignored the privacy settings attached to requests, such as Zero Data Retention (ZDR), which is supposed to guarantee that nothing is kept.
In one of the three compromises, the attackers reportedly reused a few hundred AWS keys, all sent during short windows in late October, and every request sent during those windows was reportedly stolen. Some of the stolen keys came from requests the provider had rejected before a model even ran: Moore concludes the copying happened at the service’s front door, not somewhere deep in the inference stack. In the other two cases, ZDR-flagged requests were part of the haul. The keys were then tested, sometimes within hours of being sent.
OpenRouter peels off a label
The study already has a sequel. According to an Ars Technica update dated November 21, which draws on posts published on Hacker News and X, OpenRouter reportedly wrote to its customers to warn them that some of their prompts may have been read. The router reportedly cites the study and places the potential exposure with prompts sent to MiniMax, a model provider based in China. The booby-trapped URLs were reportedly visited whether the prompts reached MiniMax directly or through three resellers: AtlasCloud, Novita and StreamLake. OpenRouter has removed the “Zero Data Retention” label from the MiniMax model endpoints.
A flag is not a vault
The experiment does not say every provider leaks. It shows that a privacy setting is a polite request made to each link in the chain, and that the guarantee is only as good as the sloppiest one. If your prompts contain keys, passwords or internal addresses, they are read in plain text at every stop, by companies you did not necessarily pick.
ZDR promised nobody would keep your prompt. It never said a word about who might be reading over its shoulder.
Sources (1)
Written with AI assistance from the sources cited above, then reviewed and approved before publication by Sébastien Soulier.


