HD Moore planted 2 million tripwires in AI routers. The data leaked.
A researcher seeded 2 million unique secrets into prompts routed through AI gateways and found sensitive credentials leaking across the chain, a direct hit on how companies buy and trust model access.

Security researcher HD Moore seeded 2 million unique canaries, including AWS keys and SSH passwords, into prompts routed through AI brokers such as OpenRouter and Vercel, and found an ecosystem that leaked highly sensitive data like a sieve. For decision-makers, the result turns AI routing from a cost and flexibility choice into an unmanaged supply chain risk that demands its own vendor review, logging rules, and board-level questions.
Security researcher HD Moore planted 2 million digital canaries inside requests flowing through a sprawling network of AI services, and the results are a warning shot for every company that pipes sensitive data through a third-party model. The number is not a metaphor. Moore assembled a collection of Amazon Web Services access keys, working credentials tied to empty AWS accounts, passwords for services such as SSH, and unique URLs, DNS names, and private git repository references. He seeded each secret into exactly one prompt sent to a single provider. Then he watched what came back.
Moore called them canaries because they mimic the role of the bird taken into coal mines: if one dies, you know the air is bad. Because each secret was unique and had never been used before, and because Moore controlled the infrastructure that would receive any request using those credentials, any hit on a canary would tell him exactly which link in the chain was responsible for the leak. What he found, according to Ars Technica's writeup of the experiment, was an ecosystem that leaked highly sensitive data like a sieve.
To understand why, you have to understand how modern AI apps are wired. Many of today's AI apps do not call Anthropic, OpenAI, or another big provider directly. They use services such as OpenRouter or Vercel that act as brokers. These AI routers, sometimes also known as AI gateways, receive large numbers of prompts sent from AI app end users and route them to an available inference engine based on criteria such as which one is charging the lowest price or which is best suited for the given task. The developer hires a router to find the most suitable model, and in doing so hands over a stream of prompts that may contain credentials, customer data, or internal code.
The exposure categories the experiment points to are the ones that keep security teams up at night: security breaches, prompt theft, and privacy violations. Those are not hypothetical labels in this context. If a prompt contains a live AWS key and that prompt is routed through a chain of intermediaries, every hop is a place where the secret can be logged, cached, resold, or simply mishandled. The canary experiment is designed to detect exactly that kind of handoff, because a canary that fires identifies not just that a leak happened but where in the chain it happened.
For context, routing layers exist because the model market is fragmented and fast moving. Historically, developers have used gateways to avoid locking into a single vendor, to fail over when one provider is down, and to chase lower prices as new models launch. The pitch is flexibility and cost control. The trade-off is that the gateway becomes a new concentration point for data. A company may have a carefully negotiated data processing agreement with one model provider, then route the same traffic through a broker that fans it out to several inference engines, each with its own logging and retention practices.
Regulators have been circling this exact gap. Privacy regimes in Europe and elsewhere put the burden on the data controller to know where personal data goes and who processes it, and cross-border transfer rules make every additional vendor a potential compliance question. The practical problem is that a router can change the downstream provider on a per-request basis, based on price or task fit, which means the answer to "where did this data go" can vary minute to minute. That is a hard story to tell a regulator, and a harder one to tell a customer after a breach.
For executives and boards, the canary experiment reframes AI procurement as a supply chain risk problem, not a tooling decision. The questions to ask are concrete: Which prompts leave our environment, which gateway sees them, what does that gateway log, how long is it retained, and can we prove the answer? If the answer is that the router decides, that is a finding. The same logic applies to any vendor that sits between your data and a model, including observability tools, agent frameworks, and internal platforms built on top of third-party APIs.
Moore's experiment does not name and shame a single culprit, and it does not need to. The lesson is structural. A 2 million canary test is a stress test of trust, and the system failed it in a way that is visible to anyone who looks. The companies that move fastest to map their own prompt supply chain, and to force disclosure from the routers in the middle, will be the ones who avoid being the case study next time. The canaries are already singing.
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.


