Security

2 Million Canary Secrets Show AI Routers and Resellers Leaking Prompts, and Three Providers Breached

Leave a comment

Ars Technica has a striking story. Researcher HD Moore seeded nearly 2 million AI prompts with unique, trackable secrets (working AWS keys tied to empty accounts, SSH passwords, unique URLs and DNS names, private repo references), one secret per prompt per provider, then watched where they surfaced. Many AI apps send prompts through brokers like OpenRouter or Vercel that route them to whatever inference provider is cheapest or best suited.

The findings: evidence that at least three AI providers were compromised by attackers siphoning the prompts they received; multiple resellers logging prompts that then reached people who used the credentials; and multiple cases of providers ignoring privacy flags like Zero Data Retention. At one breached provider, every request sent during certain late-October windows was stolen, including some rejected before a model ran, meaning attackers were copying traffic at the front door. Stolen bait keys were tested within hours, used to dig through S3 buckets and secrets managers, try to enable Bedrock models and add IAM users, and were used “tens of thousands” of times. Prompts are encrypted in transit but decrypted at each hop.

My take

Every hop in the chain reads your prompt in plain text, and a ZDR flag means nothing if the attacker is already inside. Never paste live secrets into prompts, scope your agents’ keys tight, and know exactly which providers your router can send your traffic to. I wrote up a simple canary how-to in this batch.

God Bless, and Needs Garlic!

References & Further Reading

Subscribe (RSS) Follow Security via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.