tech
2 Million Planted Secrets Reveal AI Prompt Router Leaks

A researcher sent 2 million fake secrets — AWS access keys, SSH passwords, private Git repository links and DNS names — into commercial AI services to see whether that data stayed private. It did not. According to a Nov. 20 report in Ars Technica, at least three AI providers were compromised by attackers who intercepted and reused the planted credentials, and multiple resellers in the AI inference market logged prompts despite customer settings meant to block that retention.
How did the test work?
HD Moore, chief research officer at security firm Dragos, built roughly 2 million synthetic secrets that had never been used anywhere else, Ars Technica reported. Each one — tied to an empty AWS account or a dummy SSH credential, for example — was embedded in exactly one prompt sent to a single provider. Moore then monitored the infrastructure tied to each credential, so any later access attempt pointed back to the specific service that leaked it, according to the report.
What did the canaries catch?
The experiment turned up what the Nov. 20 Ars Technica report called security breaches, prompt theft and widespread failures to protect customer data. Moore found at least three AI providers whose systems were compromised by outside attackers who then used the stolen access to pull prompts flowing through them, per the report. Separately, multiple resellers inside the AI "inference marketplace" logged prompts and passed that logged data along to people who later used the embedded credentials. Moore also documented cases in which providers ignored privacy flags such as Zero Data Retention or "no data collection" settings that customers had explicitly selected, Ars Technica reported.
"We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced," Moore wrote, according to Ars Technica. "What we found revealed security breaches, prompt theft, and widespread failures to protect customer data."
Why does a prompt get exposed at every hop?
Prompts routed through AI brokers such as OpenRouter or Vercel are encrypted only in transit, Ars Technica reported. Each AI "router" or "gateway" must decrypt the plaintext to decide which inference provider — selected for price or task fit — should handle the request. From there, a prompt can move to a pool of models hosted by different providers, or pass from one router to another, and each stop in that chain is capable of logging and storing the text, according to the report.
Why does this matter beyond one research project?
The leak sits inside the same AI infrastructure stack now under financial scrutiny for separate reasons. Anthropic shares fell this week on renewed concern about AI capital spending, HTT News reported, and the SEC is examining Nvidia's accounting practices, according to Reuters reporting cited by HTT News. Neither story involves Moore's canary data directly, but both point to an AI supply chain — financial and technical — still being pressure-tested in public.
What should businesses and users watch for next?
- Whether named routers and resellers confirm or dispute Moore's findings and disclose which providers were compromised.
- Whether any AI vendor documents how it honors Zero Data Retention settings end to end, not just at its own layer.
- Whether enterprise customers start demanding contractual logging audits before routing prompts through third-party gateways.
- Whether consumer apps that skip multi-hop AI routing, such as AlterEgo, use Moore's findings to market keeping prompts off third-party gateways.
Questions
What is an AI canary in this context?
It is a unique, trackable fake secret — such as an unused AWS key or SSH password — embedded in a single AI prompt so a researcher can see exactly where it resurfaces, according to Ars Technica's Nov. 20 report on HD Moore's research.
What is an AI router or gateway?
Services such as OpenRouter or Vercel that receive prompts from AI apps and forward them to an available inference provider based on price or task fit, decrypting the prompt at each step, per the Ars Technica report.