Unbiased headlines. Facts, not spin.
Unbiased Headlines is an independently owned news organization. We bring you the facts, not opinions, written fresh from 110+ named sources. Sources linked so you can verify it yourself.
Researcher's 2 Million Canary Prompts Show Three AI Hosting Providers Breached, Dragos' HD Moore Says

HD Moore wanted to know what happens to a secret after you paste it into an AI app. So he sent nearly two million prompts through the AI supply chain, each carrying a unique fake secret, and watched where they surfaced.
Moore, chief research officer at the security firm Dragos, published the results Thursday, Nov. 19. His summary: "security breaches, prompt theft, and widespread failures to protect customer data."
How the test worked
Moore built a set of canaries: AWS access keys tied to empty accounts, SSH passwords, unique URLs and DNS names, and private git repository references. None had ever been used. He controlled the infrastructure that would receive any attempt to use them.
Each secret went into exactly one prompt sent to exactly one provider. If it showed up somewhere it shouldn't, he could trace which link in the chain leaked it.
The study ran from July to November 2026. It targeted a layer most users never see: AI routers and gateways such as OpenRouter and Vercel. These brokers take prompts from apps and forward each to whichever inference engine fits, often the cheapest one.
Prompts are encrypted in transit. They are decrypted at each hop, because the router has to read the text to decide where to send it. Each intermediary can log the prompt and keep it for as long as its policies allow, or indefinitely.
Three hosting providers breached
The most serious finding: attackers had compromised three separate AI hosting providers. None of the three knew until Moore notified them.
In one breach, attackers reused roughly 200 of the planted AWS keys. Each key had been sent in a single request, all within a few short windows in late October. Moore wrote that every request sent during those windows was stolen.
Some of the stolen keys came from requests the provider rejected before any model ran. Moore concluded the attacker was copying requests at the provider's front door, not deeper in the inference stack.
In the other two breaches, some of the stolen requests carried zero-data-retention flags. "Privacy flags are useless if the attacker is already on the inside," Moore wrote.
The attackers moved fast. They tested keys sometimes within hours of receiving them. They searched S3 buckets, secret managers and parameter stores for more credentials. In one case they enabled a model on AWS Bedrock and added an IAM user to keep access. In total the stolen decoy keys were used in tens of thousands of AWS API calls.
No real customer accounts were involved in these calls. The keys pointed at empty accounts Moore controlled, which is what let him watch the attackers work. Moore's data puts the earliest confirmed theft in mid-July, close to the start of the experiment.
Resellers, logging and a Chinese chat room
Outright hacking was only part of it. Moore found that some resellers in the inference marketplace recorded and shared prompts through built-in logging. Some providers did not honor privacy settings such as zero data retention and "no data collection."
Moore traced roughly 2,000 inference routes. Several hundred ran through unidentified systems, and a handful sent images outside the advertised processing path. Specific examples include an undisclosed OpenRouter hop between a service called SurplusIntelligence and Anthropic, a seller adding about 1,000 hidden tokens to each prompt, and another fetching images through a shared consumer subscription.
The China finding is the one that should get attention in Washington. A canary URL appeared in a Feishu company chat 24 days after its originating prompt, and six devices in Beijing and Shanghai opened it within two minutes. Dragos said the prompt went through a U.S.-based provider with zero data retention enabled. Prompts submitted through U.S. providers also reached an unnamed Chinese AI lab, which fetched secret URLs even from requests marked zero data retention or "no data collection."
The Chinese lab has not been named. A U.S. company promising customers that data stays put, with a prompt landing in a Chinese lab anyway, is exactly the data-residency problem lawmakers say they worry about.
Limits of the findings
A canary firing proves a secret escaped and someone acted on it. It does not by itself identify which intermediary leaked it. Because the three compromised providers are unnamed, outsiders cannot independently check which companies were hit.
What providers have done
Dragos said it notified the affected providers. After notification, one router said it stopped forwarding requests to the inference marketplace implicated in the research. One breached provider published a preliminary security notice. It has not been identified, and no publication date for the notice has been given.
The mitigations the findings point to are plain. Filter secrets out of prompts before they leave your systems. Use fewer intermediaries. Avoid reseller marketplaces. Pick the upstream provider explicitly instead of letting a router choose. Seed your own canaries to monitor leaks. Where feasible, run inference on your own hardware.
The open question is which three hosting providers were breached. Until they are named, customers who sent API keys or passwords through the AI router ecosystem between July and November have no way to know whether those secrets were among the ones stolen.
Sources used for this briefing
7 links from 7 independent newsrooms (syndicated copies of the same report count once)
These are the sources it draws on, linked so you can verify.