Key facts
- Moore sent 2 million prompts carrying unique secrets through AI routers and resellers between July and October 2026.
- At least three AI providers were compromised by attackers who took the prompts they received, Moore reported.
- Attackers reused a couple hundred planted AWS keys and used stolen keys tens of thousands of times in AWS API calls.
- A Chinese AI lab fetched secret URLs from more than half the prompts sent to its models, Moore reported.
- Some resellers logged prompts and passed them to parties who used the credentials, according to Moore.
- Moore said on-premises hosting is the most effective way to prevent leakage, with direct first-tier providers next.
Security researcher HD Moore planted 2 million unique, trackable secrets in prompts sent through a network of AI routers and resellers to test whether sensitive data stays private, and reported on November 20 that it does not. Moore is chief research officer at security firm Dragos.
The experiment used AWS access keys tied to empty accounts, SSH passwords, and unique URLs, DNS names and private git repository references, each seeded into exactly one prompt sent to a single provider. Moore controlled the infrastructure behind the credentials, so any later use of them would reveal where a leak occurred.
Moore reported that at least three AI providers were compromised by attackers who used their access to take prompts they received. In one breach, attackers reused a couple hundred of the planted AWS keys, and every request sent during a few short windows in late October was stolen, including requests the provider rejected before a model ran. In the other two breaches, some stolen requests carried Zero Data Retention flags. The stolen keys were used tens of thousands of times in AWS API calls, with the parties searching S3 buckets, secrets managers and parameter stores and adding their own IAM users.
Secrets sent to two of the compromised providers were used together almost immediately from the same rented server, which Moore said indicates the same adversary was almost certainly behind both. Separately, Moore identified resellers that logged prompts and passed them on, offering low prices to draw traffic and then turning on request logging.
Moore also reported that a Chinese AI lab's systems fetched secret URLs from more than half the prompts sent to its models between July and October, including roughly two-thirds of requests marked no data collection or Zero Data Retention, with the scraper typically appearing two or three days after the request. Prompts sent to US hosting providers reselling that lab's models triggered the same scraper. A git client on a different cloud later tried to clone private repositories referenced in prompts, but only for prompts using that lab's models.
Moore said hosting models on on-premises servers is the most effective way to prevent leakage, and that receiving service directly from a first-tier provider with one contract and privacy policy is the next best option.
