Skip to main content
Researchers plant digital tripwires in AI services: a network of glowing nodes connected by lines on a dark background.

Editorial illustration for Researchers Plant 2 Million Digital Tripwires in AI Services

Researcher Plants 2M Digital Tripwires in AI Services

Researchers Plant 2 Million Digital Tripwires in AI Services

• 4 min read

HD Moore wanted to know what happens to a secret once it enters the AI supply chain, so he built 2 million of them and let them loose. Each one was a working credential: AWS access keys tied to empty accounts, SSH passwords, private git repository links, unique DNS names, none of it ever used before and all of it traceable back to infrastructure he controlled. He slipped each secret into a single prompt, sent through a single provider, then watched to see where it turned up.

The target wasn't OpenAI or Anthropic directly. It was the layer most people never think about: AI routers like OpenRouter and Vercel, which sit between developers and the big model providers, forwarding prompts to whichever engine is cheapest or best suited for the job. Developers increasingly lean on these brokers instead of picking a model themselves, which means sensitive data passes through hands nobody signed up to trust.

Moore's canaries were designed to answer a narrow question: does anything sent through that chain leak, and if so, to whom. The answer he got back pointed to a supply chain with far more holes than anyone building on top of it would want to admit.

Researchers planted 2 million digital tripwires in requests sent through a sprawling network of AI services to find out what happened to sensitive information entrusted to them. What they discovered was an ecosystem that leaked highly sensitive data like a sieve.

Why this matters

For anyone building on top of third-party AI services, this experiment should sting. Two million canary tokens leaking out of a supply chain most developers treat as a black box means the API calls, middleware, and logging layers stitched between a user prompt and a model response are a lot leakier than vendors let on. If you're a founder routing customer data through chained AI services for convenience or cost, you're trusting every link in that chain to handle sensitive fields the way you assume they do, and this research says that assumption is shaky.

Researchers in this space now have a reproducible method, planted tripwires, to actually test exposure instead of taking a vendor's privacy policy at face value. We'd treat this as a prompt to audit what data actually travels through your stack, not just what your contracts say should happen. The next thing worth watching: whether any of the services named here change their data handling in response, or whether this becomes another report that gets cited without anything downstream actually shifting.

Common Questions Answered

What were the digital tripwires that HD Moore planted in AI services?

HD Moore created 2 million working credentials including AWS access keys, SSH passwords, private git repository links, and unique DNS names that were never actually used. Each credential was traceable back to infrastructure he controlled, allowing him to monitor where these secrets ended up after being submitted through AI service providers.

What did the experiment reveal about the AI supply chain's handling of sensitive data?

The research discovered that the AI supply chain ecosystem leaked highly sensitive data at an alarming rate, functioning like a sieve. The API calls, middleware, and logging layers between user prompts and model responses proved to be far leakier than vendors publicly disclosed.

Why should developers be concerned about routing customer data through chained AI services?

Developers routing customer data through multiple chained AI services for convenience or cost are trusting every link in that chain to properly handle sensitive information, but the experiment showed this trust is misplaced. The widespread leakage of the 2 million canary tokens demonstrates that third-party AI service providers have significant security gaps in how they process and store sensitive data.

How did HD Moore track where the planted secrets ended up in the AI ecosystem?

Each of the 2 million digital tripwires was a unique, traceable credential tied to infrastructure controlled by Moore, allowing him to monitor exactly where and when these secrets appeared outside of their intended destinations. By slipping each secret into single prompts sent through various AI providers, he could identify which parts of the supply chain leaked the sensitive information.

LIVE23:21Google's Gemini Can Now Operate Mac Apps in the Background