Researcher Plants 2M Canaries in AI Supply Chain Exposing Privacy
Security researcher HD Moore seeded two million unique, trackable secrets into AI prompts across routed inference services, revealing that sensitive data routinely leaks through intermediaries. The experiment showed compromised hosting providers stole AWS keys and used them within hours, resellers logged and shared credentials, and a Chinese AI lab scraped secret URLs from most prompts, even those marked “Zero Data Retention”, demonstrating privacy flags fail when infrastructure is bre


Short answer: Security researcher HD Moore seeded two million unique, trackable secrets into AI prompts across routed inference services, revealing that sensitive data routinely leaks through intermediaries. The experiment showed compromised hosting providers stole AWS keys and used them within hours, resellers logged and shared credentials, and a Chinese AI lab scraped secret URLs from most prompts, even those marked “Zero Data Retention”, demonstrating privacy flags fail when infrastructure is breached or dishonest.
AI supply chain privacy experiment 2M canaries
Security researcher HD Moore has published results from an experiment that seeded two million unique, trackable secrets into prompts sent through a wide network of AI services, revealing that sensitive data routinely escapes the controls developers expect. Moore, who serves as chief research officer at Dragos, announced the findings Thursday. The test used Amazon Web Services access keys tied to empty accounts, SSH passwords, unique URLs, DNS names, and private git repository references. Each secret appeared in exactly one prompt directed at a single provider, and Moore controlled the infrastructure that would receive any subsequent use of those credentials. Because the secrets had never existed before, any appearance elsewhere could be traced to the specific hop in the supply chain that leaked it.
The architecture that makes this leakage possible is built on AI routers such as OpenRouter and Vercel. These gateways accept prompts from end users and forward them to whichever inference engine offers the best price or capability for the task. A prompt is encrypted while moving across the internet but must be decrypted at every intermediary so the router can read it and decide where to send it next. That means each router, each hosting provider, and each model operator in the chain sees the plaintext and may log it according to its own retention policy. Moore’s canaries showed that those logs are not staying private.
Compromised AI hosting providers data breach
The most alarming discovery involved three separate hosting providers that had been compromised by attackers. In one case the intruders harvested a couple hundred AWS keys during brief windows in late October. Every request sent during those windows was captured, including requests the provider rejected before a model ever ran, indicating the theft occurred at the front door of the service. The stolen keys were tested within hours and then used tens of thousands of times in AWS API calls. The attackers enumerated S3 buckets, secrets managers, and parameter stores, tried to enable hosted models in AWS Bedrock, and added their own IAM users to maintain persistence. Secrets sent to two of the breached providers were later used together from the same rented server, strongly suggesting a single adversary was behind both intrusions. Some of the stolen prompts had been flagged with Zero Data Retention or no-data-collection settings, demonstrating that privacy flags offer no protection when the infrastructure itself is owned.
Beyond outright breaches, Moore identified multiple resellers in the inference marketplace that were logging prompts and passing the credentials to third parties. He isolated these actors by sending the same model request through different sellers. When a secret reappeared only after traveling through a specific reseller, the responsibility was clear. These resellers attracted traffic with low prices, enabled request logging, and then monetized or shared the captured data.
Chinese AI laboratory data exposure
A separate pattern implicated a Chinese AI laboratory. Systems affiliated with that lab fetched secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked as no data collection or Zero Data Retention. The scraping typically began two or three days after the original inference call. The behavior was not limited to direct calls to the lab. Several U.S. hosting providers resell the lab’s models, and prompts routed through those U.S. providers triggered the same scraper, meaning data intended to stay in the United States flowed back to the lab anyway. In a later phase a git client on a different cloud attempted to clone the private repositories referenced in the prompts, but only for prompts that had used the lab’s models.
Other U.S. providers exhibited similar scraping. One crawled a secret URL more than two months after it appeared in a single call to a search endpoint. Another returned to scrape URLs days after each request regardless of which AI router had been used to reach it.
On-premise AI hosting defense strategy
Moore said the most reliable defense is to host models on premises, an option that remains expensive for many teams. For organizations that cannot self-host, he recommends contracting directly with a first-tier provider so there is a single agreement and a single privacy policy to evaluate, rather than a chain of intermediaries each with its own logging practices.
For engineers building AI-enabled applications, the experiment underscores that any secret placed in a prompt, API keys, database credentials, internal URLs, proprietary code snippets, should be treated as potentially public the moment it enters a routed inference pipeline. Developers should rotate any credentials that have ever been sent through third-party routers, audit their provider contracts for data-retention guarantees, and consider architectures that keep sensitive prompts inside their own network or with a single, audited vendor. The findings also suggest that compliance checkboxes such as Zero Data Retention are only as strong as the weakest link in the supply chain, and that link may be a compromised server, a dishonest reseller, or a model operator that ignores the flags entirely.
Frequently asked questions
How did the researcher track whether AI services leaked sensitive data?
HD Moore seeded two million unique, never-before-used secrets-including AWS keys, SSH passwords, and private git references-into individual prompts sent to single providers. Because each secret existed only in one prompt, any later appearance could be traced to the exact hop in the supply chain that exposed it.
Why do AI routers such as OpenRouter and Vercel create privacy risks?
Prompts are encrypted in transit but must be decrypted at every intermediary so the router can read them and choose an inference engine. Each router, hosting provider, and model operator in the chain sees the plaintext and may log it under its own retention policy, which the experiment showed are not staying private.
What happened when three hosting providers were compromised?
Attackers harvested hundreds of AWS keys during brief windows in late October, capturing every request-including rejected ones-at the front door. Stolen keys were tested within hours and used tens of thousands of times to enumerate S3 buckets, secrets managers, and IAM users. Two breaches shared a rented server, suggesting a single adversary.
Did privacy flags such as Zero Data Retention protect the canary secrets?
No. Some stolen prompts had been flagged with Zero Data Retention or no-data-collection settings, yet the credentials were still captured and abused. The flags offered no protection once the underlying infrastructure was compromised or operated by dishonest actors.
What defensive steps does Moore recommend for organizations that cannot self-host models?
Contract directly with a first-tier provider so there is a single agreement and privacy policy to evaluate, rather than a chain of intermediaries each with its own logging practices. Rotate any credentials ever sent through third-party routers, audit contracts for data-retention guarantees, and keep sensitive prompts inside your own network or with a single audited vendor.
Source: Ars Technica
Related posts

Signal Blocks Gemini and Siri From Reading Screens
Signal has disabled in-app screenshots on Android and iOS 27 to prevent OS-level AI agents like Google's Gemini and Apple's Siri from scanning private conversations. The November 20, 2026 change blocks these assistants from reading on-screen content. Users can re-enable screenshots in settings, but Signal warns this exposes messages to AI agents. President Meredith Whittaker says developers lack tools to stop AI agents from bypassing app-level privacy controls without degrading usabili

Meta Reconsiders 2030 Net Zero Goal After Record AI Emissions
Meta is re-evaluating its 2030 net-zero target after emissions surged nearly 40% to 21.7 million metric tonnes in 2025, driven by rapid AI infrastructure expansion. The company's net emissions rose 45% to 11.8 million tonnes, the steepest increase among Big Tech peers. Microsoft, Google, and Amazon reported smaller increases but have maintained their original sustainability timelines.

Hebbia Reports Gains in Financial Research With GPT-6 Astra
Hebbia tested a prerelease GPT-6 Astra in August-September 2026 and reports it improved financial research workflows. The model scored 13.6 percentage points higher than the best alternative on citing required source types, with 99.5 percent of cited excerpts relevant. It also met 97.9 percent of presentation requirements, a 14.4-point lead, ranking first on both visual polish and brief adherence. Hebbia says this could shrink analysts' overnight deck revisions to minutes of review.