Two million digital tripwires planted inside AI service requests have exposed widespread data leaks, security breaches, and prompt theft across the artificial intelligence supply chain, Ars Technica reported on Friday.
HD Moore, chief research officer at security firm Dragos, revealed the findings after orchestrating a massive test using unique, trackable secrets sent to various AI providers. The experiment uncovered that inference marketplaces, brokers, and routers routinely logged prompts and failed to honor end-user privacy flags like Zero Data Retention.
Many modern applications rely on intermediary brokers such as OpenRouter or Vercel to route user prompts to the most suitable inference engine based on price and task fitness. Although these prompts are encrypted in transit, they are decrypted at every hop within the reseller network. Each intermediary frequently logs the data, creating multiple points of exposure.

The most alarming finding involved the confirmed breach of three separate AI hosting providers. Attackers compromised these systems and reused hundreds of AWS access keys that had been seeded in the test prompts.
Ars Technica reported that some of the stolen keys originated from requests rejected by the provider before an LLM even ran, indicating that attackers were intercepting data at the provider’s front door rather than deeper in the inference stack.
Some of the compromised requests carried strict Zero Data Retention flags. As Moore explained, a privacy flag becomes useless when unauthorized actors already maintain access inside the infrastructure. Parties tied to the intrusions tested the bait keys within hours to search AWS resources, examine S3 buckets, and add their own IAM users. In total, the stolen keys were utilized tens of thousands of times in automated AWS API calls.
Tech Beat said that the experiment targeted an ecosystem that currently leaks highly sensitive data like a sieve. As applications increasingly bypass major foundational model developers in favor of dynamic routing markets, the hidden layers of routers and gateways present an enterprise risk without precedent.
“We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced. What we found revealed security breaches, prompt theft, and widespread failures to protect customer data.”
HD Moore, who founded the Metasploit Project in 2003 and served as chief research officer at Rapid7 until 2016, now co-founds and leads runZero as chief technical officer. He developed security utilities for the United States Department of Defense as a teenager and later initiated the “Month of Browser Bugs” project in 2006, which spurred a wave of browser patches and the broader Month of Bugs movement.
said HD Moore, chief research officer of Dragos, writing on Thursday.