How safe are AI API routes via resellers and routers?
Security firm Dragos sent about two million requests with fake credentials to AI providers and found compromised services and resellers reading customer prompts. Companies that use AI through third parties need to know who sits on the path of their data.

What the Dragos researchers tested
Researchers at the security firm Dragos sent roughly two million AI requests containing fake credentials to a range of providers. Over five months, from July to November 2026, they found at least three compromised providers where attackers were actively stealing user data. None of the three knew about the problem until Dragos told them. The findings were reported by The Decoder.
In the largest incident, hundreds of AWS keys were captured, some from requests the provider had rejected before processing them. Attackers tried some of the stolen keys within hours, searched S3 storage and tried to secure lasting access. One attacker also used ARTEX, an AI-assisted attack tool.
Privacy settings such as zero data retention were often ignored. A Chinese AI lab retrieved the secret test URLs from more than half of all requests, even when they went through US providers. One large US provider crawled a test URL more than two months after a single request.
Why intermediaries add risk
AI routers such as OpenRouter offer one interface to models from many providers. Requests go to a pool of providers, sometimes through several hops. According to Dragos, each hop can read the full request and the answer.
Some routers work as open marketplaces where the cheapest provider wins. Dragos reports that marketplace resellers read customer requests without exploiting any flaw. They only had to offer a model at the lowest price and switch logging on. In one case, someone tried to log in over SSH with passwords taken from intercepted requests.
OpenRouter removed three providers and withdrew the zero-data-retention approval from 125 endpoints, partly because of the findings and partly as a precaution. It also emailed affected users that requests to MiniMax models may have been read by third parties. Novita, Phala and Together have published first security reports, but other affected providers have not yet done so. The Decoder also points to an Oxford China Policy Lab analysis of Chinese proxy services, and nobody knows whether the two groups overlap.
What companies should check now
Start with an inventory of every route by which prompts leave your network. This includes developer tools, plug-ins and test projects that someone connected to a router with a private account. Each route needs a named owner.
- Which router, reseller or marketplace sits between you and the model operator, and which providers does it use behind the scenes?
- Are zero-data-retention promises backed by a contract you can read, or only by a setting in a dashboard?
- Do prompts contain passwords, keys, personal data or material covered by professional secrecy under § 203 StGB, the German criminal law provision on confidential client information?
- Who is told if a provider reports an incident, and is that contact current?
Dragos advises companies against sending requests containing credentials to intermediaries. Where personal data is involved, the questions you would ask under the GDPR apply to every link in the chain and to the model operator alike. Your data protection officer and, where you have one, the Betriebsrat (works council) should see the inventory.
Protect keys and spot misuse early
Keys that reach a compromised provider can be tested within hours, so rotation and narrow permissions limit the damage. Give each application its own key, restrict it to the services it needs and set short lifetimes where the provider allows it. Keep keys out of prompts, chat windows and code snippets that people paste into AI tools.
Dragos also recommends canary tokens, which are fake credentials whose use triggers an immediate alert. Placed in test prompts or documents, they show whether data on a given route is being read or reused. Combine them with usage monitoring and alerts for unusual activity, such as logins from unexpected locations or access to storage that the application never touches. Our overview of security and hosting lists further points to raise with providers.
Choose the data path yourself
Dragos advises companies to select the path of their AI requests themselves. That means calling the model operator or a hosting provider directly under a contract, instead of letting a marketplace pick the cheapest option. A local AI setup removes the external hop for the models that run in your own network.
souveraen.ai offers two modes of the same software: On Demand, hosted in Germany, and on-premise with Private Spark, which runs on an NVIDIA DGX Spark in the company network. The model catalogue contains local models, models hosted in Germany and external EU models with zero data retention. As with any provider, ask for the contract terms and the route behind each model before you rely on a promise.
Whatever you choose, write down the approved routes and block the rest at the firewall or through policy. A short list of permitted providers is easier to audit than a long list of tools that staff have connected on their own.


