Bait AI prompts exposed breaches at three AI providers, Dragos says

HD Moore's team seeded about two million prompts with fake credentials and links, then tracked who used them. The trail led to stolen cloud keys and prompts read by people who were never meant to see them.

Illustration of glowing envelopes flowing on cables between server racks, with two yellow canaries perched on the wires and a gloved hand pulling a gold key from one envelope. Image: AgenticFeed illustration

Attackers stole data from at least three AI hosting providers this year, and none of them knew about the breach until researchers told them, according to a study published Thursday by HD Moore, chief research officer at the security firm Dragos. His team sent about two million prompts carrying unique fake credentials and links to widely used AI services between July and November 2026, then tracked where those secrets turned up.

The study, posted on the Dragos blog, does not name the affected providers. The earliest incident dates to mid-July, Moore wrote, and one breached provider has published a preliminary security notice while another has not disclosed its incident.

Stolen keys put to work

The largest breach involved a couple hundred of the team's AWS keys, each sent in exactly one request to one provider during a few short windows in late October, according to the write-up. Every request sent during those windows was stolen, Moore wrote, including some that the provider rejected before a model ever ran. He said that shows the attacker was copying requests at the provider's front door.

The attacker tested each key, sometimes within hours, and then went looking for data in S3 buckets and AWS secret stores, the study says. It later tried to enable hosted models in AWS Bedrock and add its own IAM users. The bait keys were used for tens of thousands of AWS API calls.

On November 1, the same attacker added ARTEX, an LLM-driven penetration testing framework, though Moore wrote that it did little beyond confirming keys were still live. He flagged two user agents it sent, artex-agent and artex-judge-r121, as worth adding to detections. In two of the breaches, some stolen requests had been marked for Zero Data Retention (ZDR), a setting meant to stop providers from storing prompts.

Resellers who only had to turn on logging

The second finding required no exploit, according to Moore. Sellers on an AI inference marketplace, where the lowest-priced offer wins the request, were able to view and keep customer prompts using standard logging features. In one case, someone on a home broadband connection opened two secret URLs about a day and a half after the team sent them through a single seller's offer, then started guessing usernames at the login page.

In another case, a person working from a Mac on a home connection in France took passwords from two prompts and tried to log in as root over SSH, the study says. A separate router had been quietly forwarding some traffic to the marketplace and has since stopped, Moore wrote.

Privacy flags ignored

The third finding involved scraping rather than stolen credentials. A Chinese AI lab's systems fetched the secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked "no data collection" or ZDR, the study says. Prompts sent to U.S. hosting providers that resell the lab's models triggered the same scraper, according to Moore.

In one case, a secret link was posted into a Feishu company chat 24 days after the team sent the prompt to a U.S.-based provider with the ZDR flag. Within two minutes it was opened from desktops and phones in Beijing and Shanghai, Moore wrote. He said large U.S. providers have their own issues, citing one that kept crawling a secret URL more than two months after a single call to its search endpoint.

The team also built a tracing method for AI inference routes. Of about two thousand routes traced so far, several hundred included a system it could not identify, Moore wrote.

Advice for teams running agents

Dragos said it is working with about a dozen providers. Moore advised teams to choose upstream providers themselves instead of accepting a router's default pool, and to plant their own canary credentials in the context their agents and pipelines already send. "Keep in mind that a privacy flag only binds the first hop, and that AI agents will happily send whatever they can read, including credentials," he wrote.

The findings come from Dragos's own monitoring and have not been independently verified.

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

Additional sources: HD Moore on Infosec Exchange, “We seeded 2m canaries into popular AI services”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 30 Nov, 2026

AWS re:Invent 2026 in Las Vegas

AWS’s annual conference runs Nov. 30 to Dec. 4. Matt Garman’s CEO keynote is on Dec. 1, and the catalog lists more than 2,200 sessions.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.