Saturday, 21 November 2026, story 5 of 7
HD Moore's canary prompts trace stolen secrets through AI routers
Dragos researcher says attackers stole keys from three providers, resellers logged prompts, and privacy flags such as ZDR were ignored.

What happened
Dragos researcher HD Moore sent nearly two million prompts through AI routers, each seeded with a unique, trackable secret, and watched where they surfaced. He says attackers breached at least three providers and copied requests, some marked Zero Data Retention. Stolen bait keys were used "tens of thousands" of times in AWS API calls.
Some low-priced resellers logged prompts and passed them on, he says. A Chinese AI lab's systems fetched secret URLs from over half the prompts sent to its models between July and October, including about two-thirds of those flagged no data collection or ZDR. The article does not name the compromised providers.
What it means
A privacy flag is a request, not a control. If Moore's account holds, ZDR settings did not protect prompts once a provider's front door was compromised, and some providers did not honour such flags at all. Routers read plaintext to choose a model, so every hop is a place where a credential can be copied. Teams that paste live keys, passwords or private repository URLs into prompts should treat them as exposed and rotate them.
The limits matter. This is one researcher's experiment, seen through one article, and the compromised providers are unnamed, so teams cannot check whether their own vendor is involved. Moore's advice, as reported, is to self-host or go direct to a first-tier provider, and to pick upstream providers rather than accept default pools. Cheap routing may carry a privacy cost.
What to do: Audit which routers and resellers your apps use, pick upstream providers explicitly, and rotate any credential that has ever appeared in a prompt.
Source
Filed under Safety and security. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.