Security researchers at Dragos tested roughly two million AI requests and found that several providers allowed attackers to steal credentials without detection.
The tests ran between July and November 2026. Investigators identified at least three providers where user data was actively harvested. None of the companies knew they had been breached until Dragos contacted them.
In the largest incident, attackers collected hundreds of AWS keys. Some of these came from requests the provider had previously rejected. In other cases, thieves tested the stolen keys within hours, searched S3 storage, and attempted to set up persistent access. One attacker used ARTEX, a tool designed to automate AI-based attacks.
Marketplace resellers also read customer requests. Dragos states they did not need to exploit a technical vulnerability. Listing a model at the lowest price and enabling logging was sufficient. In one instance, a person on a residential internet connection in France attempted to log in over SSH using passwords pulled from intercepted requests.
Privacy settings such as “Zero Data Retention” were frequently ignored. Dragos notes a Chinese AI lab accessed secret test URLs in more than half of all requests, even when those requests went through US providers.
Even major US providers failed to follow their own rules. One company continued crawling a test URL for more than two months after a single request. In another case, a secret test link appeared in an internal chat on Feishu, known outside China as Lark, 24 days after the request. The link was opened within two minutes on three computers and three phones in Beijing and Shanghai.
OpenRouter removed three providers from its platform and revoked Zero Data Retention approval for 125 endpoints. This action followed the research findings and included precautionary steps. The company also emailed affected users, warning that third parties may have read requests sent to MiniMax models.
Affected providers, including Novita, Phala, and Together, have published preliminary security reports on the incident, though not all have done so yet.
Intermediaries are the weak link
AI routers like OpenRouter give users a single API for models from different companies. Requests are routed to a pool of providers, sometimes across several hops. Dragos says each hop can read the entire request and response.
Some of these routers operate as open marketplaces where the cheapest provider gets the traffic. In this case, providers including Novita, AtlasCloud, and StreamLake sat between OpenRouter and model developer MiniMax. Together runs its own infrastructure, and Phala has its roots in Web3.
Dragos warns companies against sending requests containing credentials to intermediaries. The researchers recommend that organizations control how their AI requests are routed, check their privacy settings, and plant their own canary tokens. These are fake credentials that give themselves away as soon as someone uses them.
“None of this means that you should stop using AI services. It does mean that where your prompts go is a decision, and if you don’t make it, the lowest price will make it for you”, Chief Research Officer HD Moore writes.
A recent analysis by the Oxford China Policy Lab showed how Chinese proxy services route requests to Western AI models like Claude through servers abroad. According to the analysis, those services can also read the requests in transit. It is still not clear whether the providers Dragos examined share any personnel with the Chinese proxy networks.
But the methods follow the same basic pattern. In both cases, intermediaries sit between users and the model developer, can view requests, and pass data on to third parties, often without users knowing. Data ends up with Chinese actors either way, even though users never meant to send it there, and privacy promises break down along the supply chain.



