AIToday
Large Language ModelsAI Safety & AlignmentTHE DECODERPublished: Nov 22, 2026, 22:01 JST

Dragos: 2 million AI requests expose data theft at providers

Send on LINE
Dragos: 2 million AI requests expose data theft at providers

3 Key Points

  1. What happened

    Dragos sent roughly 2 million AI requests with fake credentials from July through November 2026 and found at least three compromised providers stealing user data; OpenRouter removed three providers and revoked ZDR approval for 125 endpoints.

  2. Why it matters

    Even providers that did not know they were compromised could expose customer data, and marketplace resellers could read requests simply by logging them, so routing choices expose requests to third parties.

WHO IT HITSCompanies sending prompts containing credentials to AI intermediaries are directly exposed; security and IT teams at those organizations may need to control routing and plant canary tokens, as Dragos recommends.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Dragos's tests ran from July through November 2026, and none of the three compromised providers knew they had been breached until Dragos notified them. In the largest incident, attackers harvested hundreds of AWS keys and in some cases tested them within hours, searched S3 storage, and tried to establish persistent access; one attacker also used an AI-powered attack tool called ARTEX. Marketplace resellers did not need a vulnerability at all—listing a model at the lowest price and turning on logging was enough to read customer requests. Privacy settings like Zero Data Retention were frequently ignored: a Chinese AI lab accessed secret test URLs in more than half of all requests, even through US providers. One major US provider was still crawling a test URL more than two months after a single request. These findings sit alongside a recent Oxford China Policy Lab analysis showing Chinese proxy services routing requests to Western AI models through servers abroad and reading them in transit. Dragos says it is still not clear whether the providers it examined share personnel with those proxy networks, but the pattern is the same: intermediaries between users and model developers can view requests and pass data to third parties.

FAQ
Which providers were affected?
Dragos identified at least three compromised providers, including Novita, Phala, and Together, which have published preliminary security reports.
What did OpenRouter do in response?
OpenRouter removed three providers from its platform and revoked ZDR approval for 125 endpoints, partly due to the research and partly as a precaution.
How can companies protect themselves?
Dragos recommends controlling how AI requests are routed, checking privacy settings, and planting canary tokens—fake credentials that reveal when someone uses them.
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 articleNvidia's next CEO inherits a supervised business, investor Bill Birmingham says