Skip to content
Source: Dragos

🔓Researchers Trace 2M Test Prompts to Shady AI Routers, Some Reaching China

Your AI prompts may be read by strangers

TL;DR

Researchers sent nearly two million trackable secret prompts to AI providers. At least three were compromised, and resellers on an inference marketplace could view and keep customer prompts. If you route AI calls through third-party gateways, audit them now.

Prompts you send to AI models may not stay between you and the provider. Researchers seeded nearly two million prompts with unique, trackable secrets and followed them. At least three AI providers were compromised, and prompts sent to U.S.-based providers were passed on to Chinese AI organizations. AI routers give you one API endpoint that forwards requests to a pool of models. Every hop along that path can see your full prompt and response. Some routers are open marketplaces where the cheapest seller wins your request. If your team sends source code, credentials, or customer data through a gateway to save money, resellers you've never heard of may read it. The largest breach involved a couple hundred AWS keys stolen in late October, used for tens of thousands of API calls. Some stolen requests were marked Zero Data Retention, so privacy settings were ignored. One seeded link surfaced months later in an internal company chat in China. Whether providers publish audits, and whether router marketplaces face real accountability, is still open.

Researchers Trace 2M Test Prompts to Shady AI Routers, Some Reaching China — Dragos

Key Points

1

Researchers sent nearly two million AI prompts containing unique, trackable secrets, and at least three providers were found compromised as a result of the test.

Keep reading — it's free

Enter your email to keep reading — plus the best of AI & tech, daily. Free, forever.

Also get
or

Already a member? Sign in

ai-securityprompt-leakai-routersawsdata-privacy

Comments

Subscribe to join the conversation...

Be the first to comment