AIToday
Ars Technica AIPublished: Nov 21, 2026, 06:00 JST

Dragos's HD Moore finds 3 AI providers breached in 2M-prompt test

Send on LINE
Dragos's HD Moore finds 3 AI providers breached in 2M-prompt test

3 Key Points

  1. What happened

    Dragos's HD Moore seeded AWS keys, SSH passwords and private URLs into nearly 2 million prompts sent through AI routers; at least three providers were breached, with stolen bait keys used "tens of thousands" of times in AWS API calls.

  2. Why it matters

    The leak means sensitive credentials handed to AI services can be copied by attackers or resellers before a model even runs, so Moore says on-premises hosting or dealing directly with a first-tier provider is the safest option.

WHO IT HITSSecurity and privacy teams at companies that route prompts through AI routers or resellers are most exposed, since credentials and private URLs they embed in prompts may be logged, reused or scraped by parties they never contracted with.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The experiment was designed around a simple mechanism: each secret was planted in exactly one prompt sent to one provider, so when a credential later surfaced, Moore could trace it to a specific hop in the chain. AI routers such as OpenRouter or Vercel decrypt prompts in transit to read them and pick a model, and each party in that chain may log and store the prompt, which is how resellers and hosting providers end up holding data the end user thought was transient.

The breach findings went beyond logging. In one case every request sent during short windows in late October was stolen, including some the provider rejected before a model ran, which Moore says points to an attacker copying requests at the provider's front door. Two of the hacked providers had secrets reused almost immediately from the same rented server, suggesting a single adversary. The stolen keys were then tested against S3 buckets, secrets managers and AWS Bedrock, and "tens of thousands" of API calls followed.

Separately, Moore's canaries showed resellers offering low prices to attract prompts, then turning on request logging, and showed providers scraping secret URLs days after a request despite privacy flags. A customer who chose a US provider to keep data in the US still had prompts flow back to a Chinese lab, undercutting the assumption that provider location controls where data goes.

FAQ
How did the researcher detect the leaks?
HD Moore placed AWS keys, SSH passwords, unique URLs and other credentials into exactly one prompt each, so any later use of a secret identified the specific link in the AI chain responsible.
What does Moore recommend for keeping prompts private?
Hosting models on-premises is the most effective option but expensive; otherwise he advises receiving service directly from a first-tier provider and avoiding default provider or reseller pools.
Did privacy settings stop the scraping?
No. A Chinese AI lab's systems fetched secret URLs from roughly two-thirds of requests marked no-data-collection or Zero Data Retention, and US resellers of that lab's models triggered the same scraper.
Ars Technica AIRead Original Article
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.

Next articleAnthropic's Levent Alpöge claims Riemann Hypothesis proof