EN ▾
SNAPE REP RT Front page →
DevelopingAI safety21 Nov, 03:54 UTC

OpenRouter warns prompts sent to MiniMax models may have been read by a third party

Ars Technica illustration for its report on secrets leaking through the AI model supply chain.
Ars Technica

The notice follows a researcher's finding that unique test URLs placed in prompts were accessed 40 to 130 hours after the requests; who read them has not been established.

Read at RuntimeWire Share

Why it matters

Routers such as OpenRouter pass a user's prompt in plaintext to whichever hosting provider serves a model, so a leak at any provider exposes data the user never sent to it directly. HD Moore's canary experiment, reported by Ars Technica, seeded nearly two million prompts with trackable secrets and found breached hosts and scraping in violation of privacy terms. The OpenRouter notice is the first to name a model family. Watch for MiniMax's response and whether other routers notify customers.

Key facts

  • Test URLs were accessed 40 to 130 hours after requests
  • Notice covers prompts and outputs sent to MiniMax models
  • MiniMax is a Chinese AI lab
  • HD Moore seeded nearly two million prompts
  • Three hosting providers were found breached

Sources

More on AI safety
More from Snape ReportFront page →