HD Moore hid 2M secrets in AI prompts. Then he watched them leak.
HD Moore's 2 million canaries prove prompts do not stay where you sent them, and the AI routers sitting in the middle are the part of your stack nobody is auditing.

Researcher HD Moore planted 2 million trackable canaries inside requests sent through a sprawling network of AI services, including routers such as OpenRouter and Vercel, and found sensitive data leaking like a sieve. For decision-makers, the finding turns AI gateways into unlisted data processors and a live credential risk that most vendor paperwork does not cover.
The canaries sang. HD Moore, the researcher behind the experiment, planted 2 million digital tripwires inside requests pushed through a sprawling network of AI services, then tracked what happened to the sensitive information those services were trusted with. What came back, as Ars Technica reported, was an ecosystem that leaked highly sensitive data like a sieve. The point was not to embarrass a single vendor. Moore built the test to answer two questions at once: whether secrets can escape through the long chain of AI services that modern apps quietly depend on, and how that exposure gets used once it happens. For anyone shipping an AI feature this quarter, that is the uncomfortable part. The chain is longer than the contract.
Here is the plumbing most users never see. Many of today's AI apps do not call Anthropic or OpenAI directly. They hire an intermediary, a class of service such as OpenRouter or Vercel that functions as an AI router, sometimes called an AI gateway. The router takes the flood of prompts coming from an app's end users and sends each one to whichever inference engine fits best at that moment, judged on criteria such as the lowest price or the model best suited to the task. It is sensible engineering, and it is also a second, third and fourth data processor that the person typing the prompt never chose and, in most cases, has never heard of. Every extra hop is another place where a secret can come to rest outside the control of the company that generated it.
Moore's design is what makes the findings hard to wave away. He assembled a collection of AWS access keys, functioning credentials tied to empty AWS accounts, passwords for services such as SSH, and unique URLs, DNS names and private git repository references. None of the secrets had ever been used before. He controlled the infrastructure that would receive any request made using those credentials, so a hit would be visible and attributable rather than a vague suspicion. Then he seeded each secret into exactly one prompt sent to a single provider. That is the canary trick, named for the bird carried into coal mines: if a canary surfaces somewhere it was never sent, you know which link in the chain let it out. Two million of those tripwires, one leak path each, adds up to a map rather than an anecdote.
That map matters because enterprise data-flow diagrams are usually drawn at the vendor level. A company signs with a model provider, lists a subprocessor or two, files the paperwork and moves on. Routers break that picture. The same app can fan a single user prompt out across multiple inference engines depending on price and load, which means the model that answered yesterday's request is not guaranteed to be the model reading tomorrow's. Data residency commitments, retention policies and acceptable-use terms written against one provider do not automatically travel with the prompt. Historically, procurement and security teams have caught up with this kind of sprawl after the fact, through audits and incident reviews. Moore's experiment suggests the after-the-fact timeline is an expensive way to learn.
The exposure risk cuts in two directions. One is straight leakage: credentials or private references escaping into a system their owner never authorized. The other, flagged in the report's framing, is prompt theft and the privacy violations that travel with it. Employees paste material into AI tools constantly, and the practical question for a security team is no longer whether staff use these tools but whether anything sensitive survives the transit. A single functioning key can be enough to open a door, which is why Moore's use of empty AWS accounts and controlled receiving infrastructure was the difference between a headline and a measured result. Leaked secrets are only as dangerous as the access they unlock, but access is exactly what teams hand over when they wire an integration together.
So what does the finding actually change? It moves the router from an engineering convenience to a governance item. Start by asking, in writing, whether any AI feature in production routes through a gateway, and if so which engines receive traffic on the company's behalf. Then shrink the blast radius: credentials that are scoped, short-lived, and structurally incapable of touching production accounts, plus a clear-eyed look at whatever prompt and response logs are being retained and by whom. Contracts are the next layer. Subprocessor disclosure, notice obligations and visibility into which downstream model handled a given request all belong in the paperwork now, not after a review. None of that requires predicting the next breach. It requires accepting that the abstraction layer is part of the attack surface.
The strategic read is simpler than the technical one. Routers exist because they save money and latency, and price-based routing gets more attractive as model competition intensifies, not less. That means the middleman layer grows from here. The companies that win the next two years of AI deployment will not be the ones with the cleanest architecture diagram on the slide deck. They will be the ones who can say, without guessing, where a user's prompt actually went. Moore's 2 million canaries just showed how few organizations can answer that today.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.


