Your AI Prompt Is Not as Private as You Think
Have you ever pasted an API key, a password, or a chunk of private code into an AI prompt and just assumed nobody else would ever see it? This week gave us a clear answer, and it is not comforting. On November 19, 2026,
Have you ever pasted an API key, a password, or a chunk of private code into an AI prompt and just assumed nobody else would ever see it?
This week gave us a clear answer, and it is not comforting. On November 19, 2026, Dragos published research showing prompts sent to AI services were stolen, logged, and passed around, in some cases for months. On November 21, Together AI disclosed that an AI coding agent had leaked a credential that exposed stored customer prompts. On November 22, OpenRouter said it had removed three providers and stripped 125 endpoints of their Zero Data Retention status.
A Tripwire Inside the Prompt
Dragos did something clever. Between July and November 2026, its researchers sent about two million prompts to widely used AI services. Each carried a secret that existed nowhere else: a unique web link and fake credentials, known as canary tokens.
The canaries fired. Dragos says data from at least three AI providers was exposed, the earliest incident dating back to mid-July, and none of the three knew about their breach until the researchers told them. Some stolen requests were even marked Zero Data Retention, or ZDR, which is supposed to mean the prompt is never stored at all.
A privacy flag, as Dragos chief research officer HD Moore put it, does not help when the attacker is already inside.
Photo by Kevin Ache on Unsplash
The Agent Committed the Key
The Together AI disclosure is the part that should make every developer pause. Together AI said that during development of its public GitHub repository, called xorl, an AI coding agent committed a credential that gave access to a private Hugging Face repository.
That repository held prompt samples collected for routine software testing, from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, mainly on June 24, July 4, July 8, and July 14, 2026. Automated scanning caught the credential and the commit was reverted. But in that short window, an unauthorized party grabbed it, downloaded the samples, searched them for credentials to other services, and tested whether they worked. Five of the credentials it tested were Dragos canaries.
Together AI says it has found no evidence of additional leakage beyond this incident, and its review is still ongoing. It has made ZDR the default for new accounts and updated how it governs AI coding agents.
Still, the shape of the failure matters. No human pasted a secret into a public repo by hand. An agent with enough access to do useful work did it, and the safety net caught it only after the fact.
Every Hop Can Read the Note
Here is the simple mental model. Sending a prompt through a router or a reseller is like handing a note to a chain of couriers. Each courier can read the whole note before passing it on. Dragos found marketplace sellers who simply listed models at the lowest price, turned on request logging, and collected what flowed through. No exploit needed.
OpenRouter, which routes requests to a pool of third-party providers, said on November 22 that the findings showed no breach of OpenRouter itself. It off-boarded three providers, removed 125 endpoints from ZDR eligibility, and notified potentially affected customers. One honest caveat: Dragos has not named the three providers it found breached, OpenRouter has not named the three it removed, and neither has said whether they are the same three.
What To Do Differently
Dragos is blunt about the options. Self-hosting is the only setup where nobody else reads your prompt, and it is expensive. Going directly to a first-tier provider means one company, one contract, and one privacy policy to check. If you use a router, pick the upstream providers yourself instead of accepting the default pool.
The practical habits are simpler. Never paste a live secret into a prompt. If you ever have, rotate it. Turn on secret filtering where it exists. And consider seeding your own canary, a fake credential in the context your tools already send, so any use of it tells you a prompt leaked.
AI tools are genuinely useful, and none of this means you should stop using them. It does mean a prompt is not a private diary. It is a note passed hand to hand.
So here is my question for you: if one of your prompts from last month showed up in a stranger's hands tomorrow, what would they find in it?
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.