Dragos researchers sent about two million prompts seeded with unique credentials to AI services between July and November 2026, and say attackers copied some of those requests from inside three inference providers. According to Dragos, none of the three providers knew until the firm told them.

What Dragos says it found

The research comes from HD Moore, Chief Research Officer at Dragos. His team seeded each prompt with a canary: a URL, DNS name or AWS key that exists in only one place. Any later use of that secret shows that someone read the prompt that carried it. Dragos published the results on November 19.

The firm reports three groups of findings:

Advertisement

CyberTech Your brand belongs here. Reach the decision-makers who read CyberTech every day. Premium placements across the site and newsletter. Advertise with us
  • Data from at least three AI providers was exposed to attackers, with the earliest incident dating to mid-July.
  • Sellers on an inference marketplace could view and retain customer prompts through standard logging, with no exploit involved.
  • Several providers appeared to ignore customer privacy settings or published policies.

The three provider breaches

The largest incident involved a couple hundred planted AWS keys. Each key went to one provider in exactly one request, during a few short windows in late October, and Dragos says every request sent in those windows was stolen. Some of the stolen keys came from requests the provider had rejected before any model ran. Dragos reads that as the attacker copying traffic at the provider’s front door.

In the other two incidents, some stolen requests carried the Zero Data Retention (ZDR) flag, the setting meant to keep a prompt from being stored. Moore wrote: “A privacy flag doesn’t help when the attacker is already inside.”

Once the attackers held the planted keys, Dragos says they tested whether each key was live, sometimes within hours, then listed S3 buckets and searched Secrets Manager and Parameter Store. They also tried to enable Bedrock models, create Bedrock API keys and add their own IAM users. In total the bait keys drove tens of thousands of AWS API calls. One key seeded on October 21 was first used the next day and was still in use three weeks later.

Dragos says one adversary appears to have compromised two of the providers at once. On November 1 that attacker added ARTEX, an LLM-driven penetration testing framework, which in Dragos’s case made a handful of calls to confirm that keys already in hand still worked. The firm lists two user agents that it found in no public indicator list, artex-agent and artex-judge-r121, as worth adding to detections. Our earlier coverage of Microsoft and Check Point reports on ARTEX agents covers the framework itself.

Marketplace sellers and ignored settings

The second group of findings needed no exploit. Dragos says marketplace sellers only had to list a model at the lowest price, enable request logging and wait. In one case someone on a home broadband connection opened the secret URLs from a prompt and started guessing usernames at the login page. Prompts for the same model that went through a different seller were never touched, which Dragos reads as a leak tied to individual sellers.

The third group concerned privacy settings. Dragos reports that secret URLs were fetched by companies that were never meant to see the prompt, in some cases for requests marked as no data collection or ZDR. In one instance a link from a prompt appeared in a company chat channel 24 days after the request. Dragos is now working with about a dozen providers and says stolen keys and secret URLs were still being accessed months after the first incident.

OpenRouter’s response

OpenRouter, a gateway that routes customer requests to a pool of third-party providers, said on November 22 that it had off-boarded three providers based on its investigations. It also removed 125 endpoints from ZDR eligibility, which included providers the researchers flagged plus endpoints removed as a precaution, and notified all potentially affected customers. OpenRouter says the findings showed no breach or unauthorized access of OpenRouter itself.

Newsletter

Get the week's best tech coverage.

Free. Read by thousands of HR, tech, and business leaders.

The company lists further work underway: adapting the researchers’ testing approach to run on an ongoing basis, verifying network paths for requests forwarded to third parties, and requiring assurances before it restores ZDR eligibility to any removed endpoint.

What it means for the security leader

A prompt sent to an AI service travels through every hop between you and the model. Dragos notes that each hop can see the whole prompt and response, and that agents will send whatever they can read, credentials included. In our reading, the finding that matters most for risk owners is the scope of the ZDR setting. Dragos says a privacy flag binds only the first hop, so a procurement review that stops at the vendor’s policy page covers one link of the chain.

The same lesson appears in our coverage of what AI vendors store behind Zero Data Retention and in Wiz’s tally of AI developer platform incidents, where stolen keys also did the damage. A Salt Security survey put the share of AI-using firms that saw an AI act outside its intended behavior at 77%, so the prompt path deserves the same review as the model.

What defenders can do now

Dragos closes with options that fit different budgets. Self-hosting is the only route where no one else reads your prompts, though the hardware is costly. A direct contract with a first-tier provider gives one policy to read and the smallest exposure the firm says you can buy. Where a router is unavoidable, choose upstream providers yourself instead of taking the default pool and verify that your privacy settings change where requests go. Dragos advises staying away from reseller marketplaces.

Two controls cost little. Seed your own canaries by adding a fake credential to the context your agents and pipelines already send, and treat any use of it as an incident. Enable secret filtering where a provider offers it, with the caveat from Dragos that much sensitive prompt data does not look like an API key. In CloudTrail, add the two ARTEX user agents to detections, and keep in mind that Dragos says only a small fraction of attacker traffic announced its tooling.

Source: Dragos, “Three breaches, endless leaks: Your AI supply chain needs canaries”