← All news·2026-11-21·3 min read

Canary Test Shows AI Routers Leaking User Prompts

Researcher HD Moore sent almost two million requests with decoy tokens through AI intermediaries. According to his data, at least three providers have been hacked.

aiбезопасностьприватность

HD Moore is the chief research officer at Dragos, a security firm. He sent nearly two million requests to AI models through "routers" and resellers. These are middlemen that pass a user's request on to whichever model provider is cheapest or best suited. Each request carried one unique piece of bait, such as a working key to an empty Amazon Web Services account. Then he watched where the bait turned up. By his account, at least three AI providers had been broken into by attackers.

The reason is simple. A request travels through several companies on its way to a model, and each one opens it and reads it as plain text. That means a privacy setting is not always enough. In two of the break-ins, some stolen requests were marked "Zero Data Retention," which means the provider promised not to keep them. This matters most for people who send code, passwords or access keys through these middlemen.

Moore advises buying access directly from a large provider. Customers should also choose the providers behind a router themselves instead of accepting the default list. He says even this offers no guarantee. The report does not name the hacked services. For now, the findings rest on the researcher's own report.

Source: arstechnica.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids