Security researcher HD Moore, chief research officer at Dragos, spent five months sending AI services prompts laced with roughly two million unique "canary" secrets: fresh AWS access keys, SSH passwords, tracking URLs, DNS names and fake private-repo references. Each one went into a single prompt to a single provider, so any later use pointed straight back to where it leaked. According to coverage of his November 19 Dragos report, a meaningful number of them were read, copied or used.
Why this matters: lots of developers reach models through routers and marketplaces that forward traffic to whichever host is cheapest or fastest. Moore's results suggest the privacy promise you see on the front end, including "zero data retention" (ZDR) labels, doesn't reliably follow your prompt down that chain.
We could not read the Dragos post itself or the Ars Technica story that broke it, so the figures below come from the outlets listed under Sources. Where they disagree, we say so.
What the experiment found
| Finding | Reported by |
|---|---|
| ~2M canaries planted, roughly July–November 2026 | Cybrsecmedia ("approximately two million prompts"), AI Weekly ("nearly two million" secrets) |
| At least three AI hosting providers were breached, none named publicly | Cybrsecmedia, AI Weekly, CCTest |
| About 200 planted AWS credentials taken in several windows in late October | Cybrsecmedia, AI Weekly |
| Those keys drove tens of thousands of AWS API calls: S3 enumeration, Secrets Manager and Parameter Store searches, attempts at IAM and Amazon Bedrock | Cybrsecmedia (CCTest gives a shorter version) |
| Some secrets were copied from requests that were rejected before any model ran | Cybrsecmedia, CCTest |
| Some compromised requests carried ZDR labels | Cybrsecmedia, CCTest |
| One planted URL turned up in a Chinese company's internal chat 24 days after it was sent | Cybrsecmedia, AI Weekly |
| Stolen prompt data was used to try SSH logins against test systems | Cybrsecmedia |
According to Cybrsecmedia, Moore's team also saw attackers use ARTEX, an AI-driven pentest framework, but only to check whether stolen keys still worked. Some marketplace sellers could keep prompts just by switching on ordinary request logging. Moore traced about 2,000 request routes, and "hundreds" passed through systems he couldn't identify.
The OpenRouter and MiniMax piece
The most specific fallout so far involves MiniMax models reached through OpenRouter. AI Weekly, summarising Ars Technica, says tracking links in those prompts were fetched 40 to 130 hours after the model had already replied. That held whether MiniMax served the request itself or resellers AtlasCloud, Novita or StreamLake forwarded it.
According to explainx, which quotes an email posted to Hacker News, OpenRouter began emailing paying customers on November 21 with the subject "requests you sent to MiniMax models may have been read". Per that account:
- the email covers MiniMax prompts from July through October 2026
- OpenRouter removed the ZDR label from the affected MiniMax endpoints while it investigates
- StreamLake passed on a MiniMax statement that content moderation was on before October 22, and that MiniMax has since enabled ZDR on its side
- OpenRouter says it has asked MiniMax for its own account and had not received one when the email went out
We have not seen the email ourselves, and OpenRouter had not published a blog post as of November 21, according to explainx. MiniMax has made no public statement that we could find.
Conflicts to note. AI Weekly gives the MiniMax exposure window as September 23–October 19, which is narrower than the July–October range in the email. Explainx says the email reports that none of the AWS keys planted in the MiniMax tests were used. That probably refers to a different set of hosts than the ~200 keys Cybrsecmedia reports as stolen, but no source we read squares the two. The three breached hosting providers have not been named, and nothing we read says they were MiniMax or its resellers.
What to do
These steps come from Moore's recommendations as reported by Cybrsecmedia and CCTest. The ones marked as ours are our additions.
- Never put real credentials in prompts. Not in agent tool output, logs you paste in, or config files you ask a model to fix. If one has gone through a router or reseller, rotate it.
- If you used MiniMax models via OpenRouter between July and October, assume those prompts may have been read. Rotate any secrets they held and review what else they contained (our advice, in line with explainx's).
- Pin your upstream provider. Don't let a router pick from a default pool for sensitive work. Moore recommends dealing directly with a first-tier provider and cutting out intermediaries. Self-hosting gives the strongest protection, at a cost.
- Treat ZDR and retention promises as claims to test. They may not bind the resellers and sub-processors behind the endpoint you picked.
- Use canaries yourself. Seed unique, monitored tokens in prompts sent to third parties and alert on use.
- Audit your routing. Find out which intermediaries actually see your traffic. Moore found routes passing through systems nobody had declared.
What to watch
Watch for names: the three breached hosts, the Chinese lab CCTest says opened more than half the test links, and MiniMax's own account. And whether other routers also pull ZDR labels they can't verify.
Sources
- Cybrsecmedia — Two Million AI Prompts Expose Breaches, Data Theft and Supply Chain Failures cybrsecmedia.com
- AI Weekly — HD Moore seeds 2M AI canaries, exposes prompt leaks to China aiweekly.co
- CCTest — Two million AI canaries expose risks across the inference supply chain cctest.ai
- explainx — OpenRouter MiniMax leak email: what to do explainx.ai