1. An AI supply chain security test with two million traps
Security researcher HD Moore, chief research officer at industrial cybersecurity firm Dragos, has found that prompts sent through AI services leak far more widely than users assume. In a report published on November 19, 2026, Moore described seeding about two million trackable fake credentials and links into AI requests, then watching attackers at three compromised AI hosting providers steal and use them, exposing serious gaps in AI supply chain security.
The study ran for roughly five months starting in July 2026. Rather than probing the large model developers directly, Moore aimed at the layer of brokers, routers, resellers and hosting companies that sit between an application and the model that ultimately answers its request. That middle layer, he found, is where data slips out.
Moore is one of the best-known figures in offensive security research, having created the widely used Metasploit penetration-testing framework. His latest project applies a classic detection technique to a very modern problem: figuring out who else is reading the prompts that businesses and developers send to artificial intelligence systems every day.
2. What is a canary token?
A canary token is a decoy secret built to raise an alarm the moment someone tries to use it. It can take the form of a cloud access key, a password or a unique web address. To anyone who stumbles across it, the token looks genuine, but it is wired so that any attempt to use it alerts its owner.
The concept works much like the dye pack a bank teller slips into a bag of cash. The pack itself is harmless until a thief opens the bag, at which point it marks the stolen money and reveals what happened. In the same way, a triggered canary is proof not merely that data could have escaped, but that it actually did and that someone acted on it.
That distinction is what makes Moore's approach unusually strong. Many security studies point to theoretical weaknesses or misconfigurations that might be exploited. Canary tokens, by contrast, produce a record of real misuse, complete with timing and, in many cases, an indication of where the stolen data traveled.
3. Why AI routers and gateways became the target
Many applications do not send their prompts straight to a model developer. Instead, they route requests through intermediaries, often called AI routers or AI gateways, which accept prompts from many customers and pass each one to whichever inference provider best matches criteria such as price.
These services offer real convenience. A developer can switch between models without rewriting code, compare costs, and fall back to another provider if one is unavailable. The trade-off is that every request passes through additional hands, and each of those parties must read the prompt in plain text in order to decide where to send it.
Moore's experiment was designed to test what happens to data at those hops. By embedding unique canaries in each request, he could tell exactly which path a prompt had taken whenever one of the decoys later resurfaced.
This design also allowed him to separate different kinds of exposure. A decoy that was used to log in somewhere pointed to theft, while a link that appeared in an unexpected place showed that a prompt had been read, stored or passed along by someone who was never supposed to see it.
4. How did attackers use the stolen credentials?
Attackers who obtained the planted secrets wasted little time putting them to work. In late October 2026, roughly 200 unique Amazon Web Services credentials were exfiltrated in targeted attacks, and the thieves then tried them against Amazon's cloud platform.
The activity spanned several AWS services, including S3, Amazon's storage service, and IAM, the identity and access management system that controls who can do what inside a cloud account. Probing storage is a typical first step for an intruder hunting for valuable data, while tampering with IAM is a common way to try to secure lasting access.
Because the keys led only into a controlled environment, the attackers caused no real damage. Instead, every action they took was visible to Moore, giving him a detailed view of how quickly and methodically stolen cloud credentials are tested once they leave their owner's hands.
At least three AI hosting providers turned out to be compromised, with attackers taking customer prompt data from inside their infrastructure. That finding moves the risk beyond careless handling: outside intruders had a foothold in services that customers trusted to process their requests.
5. Why stolen cloud keys are so valuable
Cloud access keys are among the most prized items an intruder can obtain. A working key can open the door to stored files, databases, backups and computing resources, all without the attacker needing to break through a firewall or trick an employee into handing over a password.
Once inside a cloud account, an intruder can copy sensitive information, run their own workloads at the victim's expense, or create new accounts and permissions that allow them to return later even if the original key is revoked. That is why security teams treat a leaked key as an emergency and why attackers move quickly to test any credential they find before its owner notices.
The behavior Moore recorded fits that pattern closely. The intruders did not simply collect the keys and sit on them. They tried them against storage and access control systems, the two places where a real compromise would do the most immediate harm.
For businesses, the takeaway is that a credential pasted into a prompt is not a minor slip. Under the conditions Moore observed, it can become a working entry point for an attacker within a short time.
6. Logging that undercuts zero-retention promises
Not every leak involved an outside attacker. Moore found that some AI resellers and marketplace sellers were capturing customer prompts through ordinary request logging features built into their platforms.
That practice sits uneasily with zero-data-retention policies, a common promise in the AI industry under which a provider processes a request and then discards it rather than storing it. Some of the requests caught up in the compromises had been flagged for zero data retention, yet that setting did nothing to protect them once the data was in the hands of a party inside the chain.
The lesson is that a retention policy is only as strong as the weakest participant handling the data. A model developer may honor its commitment fully, but if a reseller upstream has logging switched on by default, prompts can still be stored and exposed.
7. Prompts that crossed borders
Some of the most striking findings concern where data ended up. One unique URL seeded during the study surfaced in an internal chat at a Chinese company 24 days after it was first sent in a prompt.
Moore also documented prompts originating from US-based users reaching Chinese AI organizations, despite privacy settings that should have prevented it. Users who send requests through routers frequently have no visibility into which downstream provider ultimately processes their data, or in which country it is handled.
For companies subject to data protection rules, export controls or contractual limits on where customer information may travel, that lack of visibility is a significant compliance problem as well as a security one.
The 24-day delay in the chat example is also telling. It indicates that leaked prompts are not only exploited in real time but can be retained and circulated well after the original request, long after a user might assume the data had been discarded.
8. Why the middle of the AI stack is the weak point
Much of the AI industry's effort to reassure customers has focused on how the largest model developers treat data, through enterprise agreements, security certifications and published privacy commitments. Moore's research suggests that attention has been aimed at the wrong layer.
The brokers, resellers and hosting providers that stand between applications and models are far less visible, and each one adds another chance for logging, interception or outright theft. A first-party privacy policy offers no protection against what happens at a downstream party the customer may not even know is involved.
Perhaps the most troubling aspect of the findings is the detection gap. Attackers were able to take data from at least three hosting providers over a period of months, and the companies did not catch it themselves. It took an outside researcher with a deliberately planted trail of decoys to bring the problem to light.
9. What developers and companies should take from the study
The practical message is simple: anything placed in a prompt should be treated as potentially exposed. That includes API keys, cloud credentials, passwords, internal links and customer records.
Developers routinely paste exactly that kind of material into AI tools while debugging code or automating tasks, often without a second thought. The canaries show that attackers are actively harvesting such secrets and testing them quickly, so the habit carries real risk.
Organizations relying on AI routers or resellers may also want to examine which intermediaries their requests pass through, whether logging is enabled at each step, and whether zero-retention settings are honored by every party rather than just the final model provider. Rotating any credentials that may have been shared in prompts is a sensible precaution.
Moore's study does not suggest that AI services are unusable. It does show that the convenience of routing prompts through a web of intermediaries comes with a security cost that many users have not priced in, and that the industry's protections are only as reliable as the least careful company in the chain.
