Trusence Every claim has a source
Last updated 21 November 2026 Search Türkçe
Security

Researcher's decoy secrets test where AI routers leak credentials

Any credential sent through a third-party AI router can be exposed to parties the sender never chose to trust.

A cracked vault door stands ajar beside a small brass clockwork automaton, with a key left in its lock and a magnifying glass resting on the floor.
AI-generated illustration

Security researcher HD Moore planted two million decoy credentials, including AWS keys, SSH passwords and private repository references, inside prompts routed through AI intermediaries such as OpenRouter and Vercel, which developers use to pick the cheapest or best-suited model. Each decoy was unused and sent to a single provider, so any later attempt to use one would reveal which link in the chain had passed it along. The experiment was built to show whether sensitive data entered into these brokered AI services can escape and be exploited, and the setup suggests that routers and gateways sit in the path where secrets are most exposed. The findings point to a supply chain that handles customer prompts with little visibility, which matters to anyone sending credentials or private data through third-party AI routing.

Why it matters

Developers often send prompts through routers such as OpenRouter or Vercel to pick a model by price or suitability, so those intermediaries see the full prompt contents. The experiment shows that secrets entered into these brokered services can be traced to whichever link passed them along, which means users have little visibility into where their data travels. For anyone sending credentials or private data through third-party AI routing, the practical effect is that the route a prompt takes now matters as much as the model that answers it.

Sources · 1