Researchers sent nearly two million AI prompts containing unique, trackable secrets through AI services and found evidence of data theft, security breaches and secret-URL scraping, according to researcher HD Moore.

Many AI apps use “AI routers” or “AI gateways,” such as OpenRouter or Vercel, to route prompts to inference providers. Prompts are encrypted in transit but decrypted at each step in the reseller network. Routers and other providers may log and store them under their policies.

Moore, chief research officer of security firm Dragos, seeded credentials and other unique identifiers into individual prompts. He controlled the infrastructure that would receive any requests using the credentials, allowing him to trace where the secrets surfaced.

“We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced,” Moore wrote Thursday. “What we found revealed security breaches, prompt theft, and widespread failures to protect customer data.”

The experiment uncovered breaches at three hosting providers. Attackers who compromised one provider reused “a couple hundred” of the AWS keys. Moore said every request sent during several short windows in late October was stolen, including some requests rejected before a model ran. In two breaches, some stolen requests were marked as ZDR. “A privacy flag doesn’t help when the attacker is already inside,” he wrote.

The attackers tested the stolen keys for access to AWS resources, searched S3 buckets, secrets managers and parameter stores for more credentials, and attempted to enable hosted models in AWS Bedrock. They also added their own IAM users to maintain access. The bait keys were used “tens of thousands” of times in AWS API calls.

Secrets sent to two hacked providers were used together almost immediately from the same rented server, leading Moore to conclude that the providers were almost certainly compromised by the same adversary.

Moore also found resellers that offered low prices to attract prompts, then enabled request logging. He identified them by sending requests to the same model through different sellers and checking where credentials were later reused.

The experiment also found providers scraping secret URLs included in prompts, in violation of privacy terms. Moore said a Chinese AI lab fetched URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked “no data collection” or Zero Data Retention. The scraper typically appeared two or three days after a request.

Several US hosting providers resell the lab’s models, and prompts sent through them triggered the same scraper, Moore said. He also reported that a git client on a different cloud tried to clone private repositories referenced in prompts, but only when the prompts used the lab’s models.

Moore said one US hosting provider crawled a secret URL more than two months after it was included in a single inference call to the provider’s search endpoint. Another ran a scraper that returned to URLs days after each request, regardless of which AI router was used.

Moore said hosting models on on-premises servers is the most effective way to prevent leakage, but that it is expensive. For those unable to self-host, he recommended using a first-tier provider directly, with “one contract and a privacy policy to read.” He said this cannot guarantee that no data will leak, but can reduce the chances, and advised users to choose upstream providers rather than accept default provider or reseller pools.