Two Million Booby-Trapped AI Prompts Show Where Your Secrets Actually Go Dragos chief research officer HD Moore sent roughly two million uniquely tagged prompts carrying live-but-empty AWS keys, SSH passwords and private git references through AI routers and gateways between July and November 2026, and found that three separate hosting providers had been breached, with attackers reusing "a couple hundred" of the AWS keys and generating "tens of thousands" of AWS API calls. Moore's November 19 Dragos blog write-up said every request sent during the late-October windows was stolen, including requests the provider rejected before a model ran, indicating the attacker copied traffic at the provider's front door. Some stolen requests carried a Zero Data Retention flag, which Moore noted offers no protection once an intruder is inside, and some leaks required no hacking at all because resellers on inference marketplaces logged requests to harvest credentials. Here is an old trick for catching a leaker. You tell a different version of a secret to each person you suspect, then wait to see which version turns up in the newspaper. HD Moore, the chief research officer of the security firm Dragos, ran this trick at industrial scale on the AI industry. He sent about two million prompts, each carrying its own one-of-a-kind secret, and waited to see who came knocking. Quite a lot of people did. As Dan Goodin reported for Ars Technica https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/ , the target was the middle layer of the AI business, which most users never see. Plenty of AI apps don’t send your prompt straight to OpenAI or Anthropic. They send it to a broker, an “AI router” or “AI gateway” such as OpenRouter or Vercel, which passes it on to whichever inference provider is cheapest or best suited to the job. That provider may resell someone else’s model, running on someone else’s hosting, and one router may hand off to another. The prompt is encrypted between stops, but it gets decrypted at every stop, because the router has to read your prompt to decide where to send it. Picture a letter that every post office opens to check the address, and that any of them might photocopy and keep for as long as its policies allow, which may be forever. Moore’s bait was AWS access keys that really worked but led to empty accounts, SSH passwords, unique URLs and DNS names, and references to private git repositories. None of it had been used before, and he controlled everything those credentials pointed to. According to Moore’s write-up on the Dragos blog https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens , published November 19, the study ran from July through November 2026, and the AWS accounts were fully logged. Each secret went into exactly one prompt sent to exactly one provider. So if a key turned up somewhere, he knew which link in the chain had leaked it. The front door The worst finding was that three separate hosting providers had been breached. Dragos says none of them knew until it told them, and the earliest incident dates to mid-July. At one provider the attackers reused “a couple hundred” of the AWS keys. Moore wrote: Each key was sent in exactly one request to one provider, all within a few short windows in late October, and every request we sent during those windows was stolen. Some of the stolen keys came from requests that the provider rejected before a model ever ran—this tells us that the attacker was copying requests at the provider’s front door, and not somewhere deeper in the inference stack. Whoever took the keys tested them, sometimes within hours of the bait going out. They dug through S3 buckets, secrets managers and parameter stores looking for more credentials, tried to switch on hosted models in AWS Bedrock, and added IAM users of their own so they could keep access. Altogether the bait keys were used in “tens of thousands” of AWS API calls. Keys stolen from two of the providers were used together, almost at once, from the same rented server, which points strongly to one attacker behind both breaches. In two of the breaches, some of the stolen requests carried a Zero Data Retention flag. ZDR is the customer’s way of saying “process this and then forget it.” It is a promise about what the provider will do. It says nothing about what an intruder will do, or as Moore put it: “A privacy flag doesn’t help when the attacker is already inside.” The business model Some leaks needed no hacking. On inference marketplaces the cheapest seller usually wins the request, so some resellers listed models at low prices, turned on request logging and waited. Dragos says the “deep discounts” can come from pooling unmetered consumer subscriptions, from startup credits or, in shadier venues, from billing to stolen AWS or API keys, and that some sellers take only cryptocurrency. I think the general lesson is that when someone sells you a commodity for less than it costs, you should ask what else they are selling. Sometimes the answer is you. Moore caught them by sending the same model’s traffic through different sellers. When a secret got used after passing through one seller and not another, the seller was the leak, not the model’s maker. In one case a person on a Mac in France tried SSH root logins by hand with passwords taken from the prompts, and used one of them four times in two minutes. No data collection Then there’s the polite version of the problem: providers quietly reading what they had promised not to read. Moore says an unnamed Chinese AI lab fetched the secret URLs from more than half of the prompts sent to its models between July and October. That included roughly two-thirds of the requests marked “no data collection” or ZDR, and the visits usually came two or three days later. Several US hosting providers resell the lab’s models, and prompts sent through them set off the same scraper. “A customer who picked a US provider specifically to keep their data in the U.S. didn’t get what they paid for,” Moore wrote. Later, a git client on another cloud started trying to clone the private repositories named in the prompts, but only for prompts that had gone to this lab’s models. Dragos also describes a link from a ZDR-flagged request to a US-based provider turning up 24 days later in a Feishu/Lark chat, where it was opened within two minutes from three desktops and three phones in Beijing and Shanghai. American providers had their own habits. One kept crawling a secret URL more than two months after it appeared in a single call to its search endpoint. Another sent a scraper back for URLs days after each request, whichever router the request had come through. The mechanism here is about contracts. A privacy flag is a term in your agreement with whoever you hand the prompt to. That company passes the prompt down a chain of other companies, and none of them signed anything with you. Dragos’s phrase is that “a privacy flag only binds the first hop.” In a supply chain with a dozen hops, the first one is where your protection ends. Disclosure, sort of Dragos names none of the breached providers, the resellers or the Chinese lab. It says the provider behind the mid-July incident first blamed a misconfiguration in its internal tooling, then blamed a partner’s data processing, then found that its own internal credentials had been compromised, and still hasn’t disclosed the incident publicly. One of the providers breached in late October has published a preliminary security notice. Dragos says it is working with about a dozen providers, and that the stolen keys and URLs are still being accessed. Moore’s advice runs from expensive to merely tedious. The safest option is to run models on your own hardware: Dragos lists a 512GB Mac Studio at $22,299 and an eight-GPU B300 server at $570,000 or more. Next best is to buy directly from a first-tier provider, so that you have “one contract and a privacy policy to read.” Even that comes with no guarantee. If you use a router, pick the upstream providers yourself instead of accepting the default pool, and plant a few canaries of your own. So after all the brokers, marketplaces and gateways, the best privacy feature on offer in AI is knowing which company you are dealing with.