HD Moore, director of research at security company Dragos, announced Thursday that it had sent nearly two million requests to AI services with traceable secrets. It reports thefts from three compromised hosts, but also reuse of information despite the confidentiality settings chosen.
URLs consulted despite confidentiality rules
A confidential address consulted more than two months after a single inference call: this is one of the behaviors that HD Moore says he observed among an American host. Another had an extraction tool return to received addresses several days after each request, regardless of the router used.
The case of a Chinese AI laboratory covers a period from July to October. According to HD Moore, its systems accessed secret addresses contained in more than half of the requests made to its models. This proportion reached around two thirds for those accompanied by an instruction prohibiting collection or a Zero Data Retention marking, known as ZDR. Consultations generally took place after two or three days.
Several American hosts market models from this laboratory. According to the researcher, requests passing through these providers triggered the same extractor and returned to the Chinese laboratory. It also reports that a git client, hosted in another cloud, then tried to copy the private repositories mentioned in the requests. These attempts exclusively concerned those using laboratory models.
AWS keys stolen from hosts and resellers
At one of the hacked providers, a few hundred AWS keys were reused by the attackers, says HD Moore. All experimental requests sent during brief windows in late October were allegedly stolen, including some rejected before any model execution. For the researcher, this last observation places the copy at the entrance to the service, and not in the internal layers of processing.
In the other two compromises, requests marked ZDR were among the stolen elements, he reports. People linked to the attacks tested the keys, sometimes after just a few hours, to determine their rights. They looked for alternate credentials in S3 spaces, secret managers, and settings stores, attempted to activate templates through AWS Bedrock, and added their own IAM users to maintain access. According to HD Moore, tens of thousands of calls to AWS programming interfaces were made using the stolen keys.
The evening letterFree · 7 p.m.
Are you following the race for AI models?
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
From a single rented server, secrets transmitted to two compromised hosts were used jointly, almost immediately. These facts suggest that both providers were almost certainly hacked by the same adversary.
Other reuses involve the resellers themselves: some attract requests with low prices, then activate registration. Identifiers from these recordings were used by third parties. To isolate their liability from that of the model supplier, HD Moore ran requests for the same model from different vendors and compared the observed reuse.
Intermediaries can read each request
The system was based on previously unused secrets: keys and identifiers giving access to empty AWS accounts, passwords in particular for SSH, unique web addresses, DNS names and references to private git repositories. HD Moore controlled the infrastructure receiving the connections made with these baits. Each secret only appeared in a request to a single provider. These markers are nicknamed “canaries”, in reference to the birds used in mines, because they are used here to identify the links responsible for a leak.
The targeted services are part of a chain where applications pass through intermediaries such as OpenRouter or Vercel. Rather than directly approaching OpenAI, Anthropic, or another large vendor, a developer can delegate model selection to them. These routers direct requests in particular according to the price or the suitability of the service for the task.
Encryption during transport does not prevent successive actors from reading the content: it is decrypted at each step. The router examines the text to select an inference service, then can direct it to a set of models operated by different hosts. A request can even go from Vercel to OpenRouter. Each intermediary is likely to record its content and keep it according to its own rules, possibly without expiry.
Moore recommends choosing each vendor
To limit leaks, HD Moore favors models hosted on the servers of the user organization. This formula remains expensive. When self-hosting is financially unattainable, he instead recommends a direct relationship with a large provider with a searchable contract and privacy policy.
This direct relationship does not constitute a guarantee of watertightness, specifies the researcher, but according to him reduces the risks. He also advises explicitly selecting suppliers upstream, rather than keeping the sets of providers or resellers offered by default.

