A new study by security researcher HD Moore, published on the Dragos blog on November 19, found that prompts sent to popular AI tools often pass through more companies than most users expect.
From July to November, Moore’s team sent about two million prompts to popular AI services. Each prompt included a unique bait secret, like an empty AWS key, a password, or a hidden link, so the researchers could track when someone accessed the data.
The experiment found active breaches at three hosting providers that run AI models for other businesses. Attackers were quietly copying incoming requests at all three companies, and none of the providers were aware until Dragos notified them. In the biggest case, attackers stole a few hundred AWS keys, tested some within hours, and made tens of thousands of API calls.
Beginning November 1, the intruders also used ARTEX, an AI-driven penetration testing tool recently connected to the cyberattack on the European Commission. Since the bait was mixed with real traffic, Dragos said that real customer prompts were almost certainly copied too. The company did not name the affected providers, but one has released a preliminary security notice.
These leaks mostly happen because of how modern AI apps handle traffic. Instead of connecting straight to a model, many apps and agents send requests through routers to the cheapest provider. Each stop along the way can see the full prompt. On open model marketplaces, Moore found that a seller just needed to offer low prices and turn on request logging. In one example, someone in France tried to use stolen passwords to log in as root over SSH.
Even strict privacy settings did not prevent exposure. A Chinese AI lab accessed hidden links in more than half of the prompts sent to its models, including about two-thirds that were marked for no data collection or Zero Data Retention. The same thing happened when prompts went through U.S. resellers, with one link showing up in a Chinese company chat 24 days later. As Moore said, a privacy flag only protects the first hop.
Security experts recommend never putting passwords, API keys, or private links in AI prompts, especially when using autonomous agents. Getting models directly from their developers can also lower the risk, as shown by a recent provider breach that exposed Notion AI chats earlier this month.

