Canary study finds AI providers and resellers exposed prompts and keys
A Dragos research team led by HD Moore reports sending about two million prompts to widely used AI services between July and November 2026. Each prompt contained unique, trackable URLs and decoy credentials ('canaries'). Dragos says data from at least three AI hosting providers was exposed: attackers used the seeded AWS keys to make tens of thousands of AWS API calls, and some stolen requests had been marked zero data retention (ZDR). The team also found that sellers on an AI inference marketplace could log and keep customer prompts, and that prompts sent to U.S. providers reached Chinese AI organizations. Dragos says the leaks followed specific sellers rather than whole marketplaces.
Why it's notable: A provider's privacy setting or ZDR label did not reliably keep prompts private in this test, especially when traffic went through resellers or routing services. Dragos recommends self-hosting where possible or contracting directly with a first-tier provider rather than through resellers, keeping secrets out of prompts, and planting decoy credentials to detect leaks; it notes that a privacy flag only binds the first hop.