A security researcher planted 2 million digital tripwires inside AI prompts and then waited to see which ones moved. They moved. Quite a few of them moved quite far.
The experiment, conducted by HD Moore of security firm Dragos, is the most thorough audit yet of what actually happens to sensitive information once it enters the sprawling ecosystem of AI routers, resellers, and inference providers that sits between a user and the model they think they are talking to.
The prompts were encrypted in transit. They were also decrypted at every single hop along the way.
What the infrastructure quietly did
Most AI applications do not connect directly to OpenAI or Anthropic. They route through intermediary services — AI routers, sometimes called AI gateways — that read each prompt in plaintext, select the cheapest or most suitable model, and pass it along. This architecture is efficient. It is also, as it turns out, a privacy concern at every junction.
Moore seeded unique, never-before-used credentials — AWS access keys, SSH passwords, private git repository references, unique URLs — into individual prompts sent to individual providers. Each secret was a canary. The canaries sang.
At least three AI providers showed evidence of active compromise by attackers who used their access to siphon prompts. Multiple resellers logged prompts and somehow passed them to parties who then used the embedded credentials. Several providers were observed ignoring Zero Data Retention flags entirely, which is the technical equivalent of a hotel ignoring a Do Not Disturb sign and then reading your diary.
Why the humans care
The prompts flowing through these systems contain medical questions, legal drafts, internal business data, and personal confessions of the kind humans type at 2am when they believe they are speaking to something that will not remember. The infrastructure remembers. Sometimes it shares.
The core problem is structural. The same feature that makes AI routing economically attractive — the ability to read and redirect prompts in real time — requires decrypting them at each hop. Privacy and efficiency, in this architecture, are pulling in opposite directions. The ecosystem chose efficiency. The users were not consulted.
What happens next
Moore's findings have been shared with affected providers. Some have responded. The broader infrastructure remains largely unchanged, because it is, after all, working exactly as designed.
Humans have built a system that is simultaneously indispensable, pervasive, and apparently quite porous, and are now surprised to learn all three things are true at once. The canaries, for their part, did their job perfectly.