Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

When I first heard that a researcher had slipped two‑million digital canaries into the AI supply chain, my gut clenched. HD Moore of Dragos didn’t just poke around; he flooded the ecosystem with trackable secrets and watched them spill out of every crack.

Nov 21, 2026 - 18:06
0 4
Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

When I first heard that a researcher had slipped two‑million digital canaries into the AI supply chain, my gut clenched. We’ve been hearing the same song about “low‑cost routing” and “best‑in‑class privacy” from AI brokers for years, but this experiment actually proved the hype is a leaky bucket. HD Moore of Dragos didn’t just poke around; he flooded the ecosystem with trackable secrets and watched them spill out of every crack. The result? A stark reminder that the AI “router” model—where a third‑party broker reads your plaintext prompt, picks the cheapest or most suitable model, and hands it off—has become a privacy disaster zone.

Why the AI router model is a privacy nightmare

The article makes it clear that many modern AI apps rely on services like OpenRouter or Vercel as brokers. These “AI routers” sit between the end‑user and the inference engine, inspecting each prompt in plaintext to decide which model to invoke. That inspection step alone defeats any claim of end‑to‑end encryption. Even though the traffic is encrypted in transit, the moment the prompt reaches the router it is decrypted, read, and often logged.

What’s worse, the router can forward the same request to a pool of models that may be hosted across several providers. In practice this means a single prompt can pass through multiple hands—each with its own logging policies, some of which keep data indefinitely. The experiment showed that even when a request is marked with privacy flags like “Zero Data Retention,” the data can still be harvested downstream.

The canary experiment: a stress test of the whole chain

Moore’s method was brutally simple: he generated a set of fresh AWS access keys, empty‑account credentials, SSH passwords, unique URLs, DNS names, and private git repo references—none of which had ever been used. He then embedded each secret into a single prompt and sent it to a single provider. By doing this at scale—nearly two million prompts—he created a massive, traceable trail of “canaries” that would light up any point in the chain that mishandled the data.

When a secret showed up somewhere it shouldn’t have, Moore could pinpoint the exact hop that leaked it. This approach turned the abstract notion of “data leakage” into a concrete, measurable event. The findings were sobering: the chain is riddled with breaches, prompt theft, and outright scraping of secret URLs.

Three hosting providers compromised: what went wrong?

The most alarming part of the report is the breach of three separate hosting providers. In each case, the attacker stole the canary credentials the moment they arrived at the provider’s front door—sometimes even before a model was invoked. Moore notes that the stolen keys were reused “a couple hundred” times, and the attacker quickly tested them across S3 buckets, Secrets Manager, and Parameter Store, looking for more loot.

One of the compromised providers even marked some requests with a “ZDR” privacy flag, yet the attacker still accessed the data. This tells us that privacy flags are meaningless if the adversary is already inside the provider’s network. The stolen keys were used “tens of thousands” of times in AWS API calls, indicating a rapid, automated exploitation once the breach was confirmed.

Resellers turned rogue: the low‑price lure

Moore also uncovered that some of the leaks originated not from the model providers but from the resellers themselves. By offering rock‑bottom prices, these brokers attracted traffic, then turned on request logging and harvested the secrets. The researcher identified the culprits by sending identical prompts through different sellers; when a secret leaked only via one path, the guilty reseller was exposed.

This business model is a classic case of “you get what you pay for.” The promise of cheap AI services masks a hidden cost: the loss of any confidentiality you thought you were preserving. For a founder who’s built a hosting platform from the ground up, it’s a red flag that the cheapest route is often the most dangerous.

Scraping secret URLs: the Chinese lab and US providers

Perhaps the most unsettling discovery was that a Chinese AI lab was actively scraping secret URLs embedded in prompts. Moore reports that the lab fetched URLs from more than half of the prompts sent to its models between July and October, even when the prompts were flagged as “no data collection.” The scraper typically appeared two or three days after the request, showing a systematic effort to harvest data.

The problem didn’t stop at the lab’s own service. Several US hosting providers that resell the lab’s models also triggered the same scraper, meaning that a customer who deliberately chose a US‑based provider for data residency got exactly the opposite. One US provider even continued crawling a secret URL more than two months after a single inference call, and another ran a scraper that fetched URLs days after each request, regardless of the router used.

What this means for independent hosting providers

From a business‑risk perspective, the findings are a wake‑up call. If you’re running an AI‑enabled SaaS product and you rely on third‑party routers or resellers, you’re effectively handing over your customers’ secrets to an opaque chain of actors. The risk isn’t just theoretical; the experiment shows that even a single compromised router can expose every prompt that passes through it.

For independent hosting providers, the lesson is clear: you either need to own the entire stack or you need iron‑clad contracts that limit data exposure. The research suggests that the “best practice” of using a pool of models from multiple providers is a security nightmare. Each additional hop multiplies the attack surface, and the logging policies of each participant are often unknown.

Actionable steps for founders and operators

So what can you do right now? First, audit every point in your AI workflow. Identify if you’re using an AI router, a reseller, or a direct API to a first‑tier provider. If you’re using a router, demand transparency: how long do they retain prompts, do they log them, and can they prove they never scrape URLs?

Second, consider moving critical workloads to on‑premises inference where you control the hardware and the data path. Moore acknowledges that this is expensive, but for high‑value secrets it may be the only viable risk mitigation. If on‑prem isn’t feasible, the next best option is to contract directly with a first‑tier provider that offers a single, clear privacy policy—no multi‑seller pools, no hidden logging.

Third, implement secret‑management hygiene at the application layer. Never embed raw credentials in prompts; use tokenization or one‑time secrets that become useless if leaked. And finally, monitor for anomalous activity on your cloud accounts—unexpected API calls, new IAM users, or sudden access to S3 buckets can be early indicators of a breach similar to the one Moore documented.

In short, the canary experiment has ripped the veil off the AI supply chain. The industry’s hype about “cheapest model” and “best‑in‑class privacy” is a mirage when the underlying architecture is built on routers that read and log everything. As a founder who’s spent a decade wrestling with production servers, I can say this: if you can’t see the data path, you can’t trust the service. The safest route is either full control or a single, transparent contract with a reputable provider. Anything else is just inviting the next wave of data leaks.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: Ars Technica; arstechnica.com; Global1.News (21 November 2026).

By Allan Ali, Global1.News

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Wow Wow 0
Sad Sad 0
Angry Angry 0
Allan Ali

Publisher of Global1.News. Automation architect, systems builder, and the guy making sure the truth gets published.

Comments (0)

User