November 21, 2026
A-researcher-introduces-2-million-canaries-into-the-AI-​​supply.jpg

The findings are surprising. They included:

  • Evidence that at least three AI providers were compromised by attackers who used their access to divert the instructions they received
  • Multiple resellers in the inference market who recorded the prompts and somehow passed them on to the people who used the credentials.
  • Multiple cases of AI providers not following end-user privacy settings embedded in flags that specify terms such as Zero Data Retention (ZDR) or “no data collection”

“We sent nearly two million AI messages containing unique and traceable secrets to AI providers, then watched where those secrets appeared,” Moore, research director at security firm Dragos, wrote on Thursday. “What we found revealed security breaches, rapid thefts and widespread failures to protect customer data.”

Although indications are encrypted in transit, they are decrypted at each hop within the large network of inference resellers. The message is first sent to an AI router, which reads the plaintext it contains to determine the most appropriate inference providers to handle the request. Often the message is sent to a group of available models, sometimes running within different hosting providers. Directions can also pass from one AI router (such as Vercel) to another (for example, OpenRouter). Each of these players is likely to record the message and store it for as long as their policies require (if not indefinitely).

The wild west in the age of AI

The most worrying finding was the non-compliance of three different hosting providers. The attackers who compromised one of the providers reused “a couple hundred” AWS keys. Moore wrote:

Every key was sent in exactly one request to a provider, all within a few short periods at the end of October, and every request we sent during those periods was stolen. Some of the stolen keys came from requests that the provider rejected before a model was run; this tells us that the attacker was copying requests to the provider’s front door and not somewhere deeper in the inference stack. In the other two breaches, some of the stolen applications were marked as ZDR. A privacy flag doesn’t help when the attacker is already inside.

The parties linked to the hack of these three providers tested each key, in some cases within a few hours of being planted, to see what type of resources it gave access to. The parties combed through S3 buckets, secret managers, and parameter stores for even more credentials. They also tried enabling models hosted on AWS Bedrock and added their own IAM users to maintain access. In total, the stolen bait keys were used “tens of thousands” of times in AWS API calls.

Avatar photo
Written by

Jhon Smith

Info Vitalis is a content writer specializing in creating clear, concise, and engaging articles. With experience in health, lifestyle, technology, and current events, Info Vitalis aims to provide readers with useful and easy-to-understand information.

Leave a Reply

Your email address will not be published. Required fields are marked *