AI router password leaks, Digimart Global: Prompts sent to AI services often reach more companies than users expect, according to a study that security researcher HD Moore published on November 19 on the blog of security firm Dragos. From July through November, his team sent about two million prompts to widely used AI services, each containing a unique bait secret such as a real but empty AWS key, a password or a hidden link.
The study found that attackers were copying requests at three hosting providers that run AI models for other companies. None of the providers knew about the activity until Dragos contacted them.
In the largest case, a couple hundred AWS keys were stolen, some were tested within hours, and they were then used for tens of thousands of API calls. Dragos writes that the bait prompts were processed alongside real customer requests and were “almost certainly copied too.” The report does not name the providers, although one has already published a preliminary security notice.
Many apps and AI agents do not connect to a model directly. They send requests through routers that forward each one to the cheapest available provider, sometimes across several hops, and every hop can see the full prompt.
Moore says that on open marketplaces, a seller only had to list a model at a low price and turn on request logging. In one case described in the study, a person in France tried to log in as root over SSH with the stolen passwords, typing each attempt by hand.
The study also reports that Zero Data Retention, the strictest privacy setting, which indicates that nothing should be stored, did not reliably hold. According to the findings, a Chinese AI lab fetched the hidden links from more than half of the prompts sent to its models, including roughly two-thirds of requests marked as no data collection or ZDR.
Similar results were reported for prompts sent to U.S. providers that resell the lab’s models.
In one case, a link appeared in a company chat in China 24 days after the request. Moore concluded that “a privacy flag only binds the first hop.”
The report’s practical advice is never to paste passwords, API keys or private access links into a prompt, even for a quick debugging session, and to be especially careful with AI agents that read files and inboxes on their own. Where possible, users should access a model directly from its developer rather than through apps that bundle many models at low prices.
A separate breach at a background provider exposed Notion AI chats earlier this month, according to the same report.














