Research

Dragos Sent Two Million Booby-Trapped Prompts to AI Providers. Three Were Breached and Did Not Know. OpenRouter Dropped Three and Pulled 125 Endpoints From Zero Retention. Treat Every Prompt as Read

Dragos Sent Two Million Booby-Trapped Prompts to AI Providers. Three Were Breached and Did Not Know. OpenRouter Dropped Three and Pulled 125 Endpoints From Zero Retention. Treat Every Prompt as Read

If a real secret has ever gone into a prompt sent to a hosted model (an AWS key, a database password, a token pasted into an agent's context), rotate it now, whichever provider you used. Then check your inbox for a notice from OpenRouter, look at which upstream providers your router is allowed to send requests to, and plant fake credentials in your agents' context so the next theft trips an alarm instead of going unseen.

The reason is research Dragos published on Thursday, November 19, 2026. Dragos "sent nearly two million AI prompts containing unique, trackable secrets to AI providers" between July and November, and watched which of those secrets were later used. Three providers had been breached, and, in Dragos's words, "none of the three providers knew about their breach until we notified them."

On Sunday, November 22, OpenRouter said it had "off-boarded three providers from OpenRouter, based on the results of our investigations," "removed 125 endpoints from Zero Data Retention (ZDR) eligibility," and "notified all potentially affected customers." OpenRouter also said the findings "did not show any breach or unauthorized access of OpenRouter." The leaks were at providers behind the router, not in the router.

What Dragos found

Each seeded prompt carried its own unique URLs and credentials, so any use of one points back to the exact request it was sent in. Three separate patterns showed up.

  • Providers that were broken into. "The earliest incident dates back to the middle of July." In the largest case a couple of hundred AWS keys were stolen in late October, and "each key was sent in exactly one request to one provider." Some were taken from requests the provider turned away: "Some of the stolen keys came from requests that the provider rejected before a model ever ran." In the other two breaches, "some of the stolen requests were marked as ZDR," the zero-data-retention setting customers pick precisely so their prompts are not kept.
  • What the thief did with the keys. Dragos says the keys were tested within hours. The attacker listed S3 buckets, searched Secrets Manager and Parameter Store, tried Bedrock and tried to create IAM users, in what Dragos counts as tens of thousands of AWS API calls. The tooling moved from a script posing as Chrome on Windows, to Boto3 in OrbStack, to the AWS CLI on Kali Linux. "One adversary appears to have compromised two of these providers at the same time." On November 1 the same actor used ARTEX, the free AI-driven penetration testing framework, only to confirm the keys were live, and Dragos records the user agents artex-agent and artex-judge-r121.
  • Marketplace sellers who did not need to break in. On reseller marketplaces a seller "only had to list a model at the lowest price, enable request logging, and wait." Dragos traced some of these sellers to home broadband lines, and passwords from two prompts were used from a home connection in France. Its conclusion: "the leak follows the seller and not the marketplace as a whole."

Dragos also reports that an unnamed Chinese AI lab fetched secret links from more than half of the seeded prompts sent to its models, including requests marked no-data-collection. We read that part of the Dragos page only through a summary because our fetch of the page was cut off, so check the original before you quote it.

Dragos's own one-line summary of the problem: "A privacy flag doesn't help when the attacker is already inside."

What is not known

  • Which providers. Dragos does not name the three breached providers. OpenRouter does not name the three it removed. OpenRouter's post links incident notices from Novita, Phala and Together; we did not read those notices, so we are not saying which, if any, of them are the three.
  • Indicators. Neither primary publishes an IP address, domain or file hash. The only technical markers in what we read are the two ARTEX user agents and the tool chain above. We added nothing to our feed for this story, because there is nothing in it to block.
  • How many customers. OpenRouter says it notified all potentially affected customers. It does not give a number.

What to do

  1. Rotate anything that went into a prompt. That includes secrets an agent read from a file, an environment variable or a tool result and then sent to a model as context. If you are not sure, assume it did.
  2. Look in AWS CloudTrail for the actions Dragos describes, coming from places you do not recognize: ListBuckets, Secrets Manager and Parameter Store reads, Bedrock calls, CreateUser and CreateAccessKey. Also search for the user agents artex-agent and artex-judge-r121.
  3. If you use OpenRouter, read any notice it sent you, then open your provider routing settings and choose the upstream providers yourself instead of letting the cheapest one win. Dragos's advice for routers is to verify that a privacy setting actually changes where your requests go.
  4. Stay off low-price reseller listings for anything that carries real data. Dragos's ranking of the safer options: run the model yourself, which is the only setup where no third party reads your prompts, or go to a first-tier provider directly and accept that you are trusting one company's policy.
  5. Plant tripwires. Put fake credentials in your agent and pipeline context (free Canarytokens work) and treat any use of one as an incident. That is the method Dragos used to find all of this, and it costs nothing.
  6. Turn on secret scrubbing where your provider or gateway offers it. It catches key-shaped strings. It does not catch a customer record or a password typed in plain words.

Am I affected?

You are affected if you sent prompts containing real secrets through a hosted model between mid-July and November 2026, and especially if they went through a router or marketplace that could pick the provider for you. The quickest check: search your agent logs, gateway logs or prompt history for strings that look like AWS access key IDs (they begin with AKIA or ASIA) and other live tokens. Every one you find should be rotated, whether or not you received a notice.

How we know, and what our edge saw

We read the Dragos post and the OpenRouter announcement directly. The Dragos page was cut off when we fetched it, at about 40,000 of 58,000 characters, so the later sections were read only as a summary and are labeled that way above. The Hacker News article that brought this to our attention also reports details from the provider notices that we did not check.

Our edge has not seen the ARTEX user agents: 0 matches in 778,582 edge blocks, 0 in 203,850 edge canary hits and 0 in 7,664 origin honeypot captures since July 1. That only covers our own sites, and this attack went after cloud accounts using stolen keys, not websites, so it is not evidence either way.

We have written about ARTEX twice before. It is free, it is in several unrelated hands, and the people using it are now using it to check whether stolen credentials work.

Read next: AI Agents Stole AI Keys From Exposed LiteLLM, Dify and One API Servers, Then Used Them to Pay for the Hunt, and on ARTEX itself, Off the Rack: A Free AI Attack Framework Is Now in Three Unrelated Hands.

Sources

  • Dragos, "Three breaches, endless leaks: Your AI supply chain needs canaries," November 19, 2026 (HD Moore, Chief Research Officer): https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
  • OpenRouter, "Data Processing Incidents on Inference Providers," November 22, 2026: https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/
  • The Hacker News, "OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms," November 22, 2026: https://thehackernews.com/2026/11/openrouter-drops-3-providers-after.html
  • Provider notices linked by OpenRouter (Novita, Phala, Together): not read for this post.

Every indicator in this post is in the feed. Free.

1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.

Get your free key → analytics.dugganusa.com/stix/register