AIToday
AI Safety & AlignmentLobsters AIPublished: Nov 22, 2026, 13:00 JST

HD Moore plants 2m canaries: three AI providers breached

Send on LINE
HD Moore plants 2m canaries: three AI providers breached

3 Key Points

  1. What happened

    HD Moore of Dragos seeded nearly 2 million prompts with unique, trackable secrets; at least three AI providers were compromised, and attackers reused stolen AWS keys tens of thousands of times.

  2. Why it matters

    The same adversary almost certainly breached two of those providers, based on secrets from both being used together almost immediately from the same rented server.

WHO IT HITSDevelopers and companies routing prompts through AI routers or resellers may have their cloud credentials and private code references exposed, and security teams will need to audit which intermediaries handle their data.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The experiment exploited a structural feature of modern AI apps: many developers do not connect directly to Anthropic, OpenAI, or another big provider, but hire an AI router to find the cheapest or best-suited model. Each hop in that reseller network decrypts the prompt to inspect it, and each player is likely to log and store it for as long as its policies call for, if not indefinitely.

The findings distinguish between two kinds of failure. In the most serious case, attackers compromised providers outright and copied requests at the front door, before any model ran. In other cases, resellers themselves logged prompts and passed them along to people who used the credentials, a pattern Moore identified by sending the same model through different sellers.

A separate concern involved scraping of secret URLs embedded in prompts, with one Chinese AI lab's systems fetching those URLs from more than half of the prompts sent to its models between July and October. Prompts sent through several US hosting providers that resell this lab's models triggered the same scraper, suggesting data did not stay in the US as customers may have expected.

FAQ
How did the researchers find the leaks?
HD Moore planted AWS access keys, SSH passwords, unique URLs, DNS names, and private git references into individual prompts sent to single providers. The secrets never appeared elsewhere, letting him identify which link in the chain leaked them.
What did OpenRouter tell customers?
OpenRouter emailed customers that some of their prompts may have been read, citing the research and a report that the exposure involved prompts sent to MiniMax. OpenRouter also removed a "Zero Data Retention" label from MiniMax-model endpoints.
What does Moore recommend to prevent leakage?
Moore says the most effective option is hosting models on-premises. For those who cannot afford that, he recommends contracting directly with a first-tier provider and avoiding default provider or reseller pools.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleMeta ads AI connector lands in AWS Quick