Security firm Dragos says attackers got inside at least three AI inference providers this year and took prompts that customers assumed were private. The evidence comes from a five-month bait operation led by chief research officer HD Moore, laid out in the Dragos research post of November 19, 2026. None of the affected providers, which Dragos does not name, had spotted the intrusions before being told.
Quick Take
- The method: From July to November, Dragos hid one-off fake keys and links in roughly two million prompts and logged where they resurfaced.
- Breaches: Intruders copied incoming traffic at three hosting companies, in one case even requests the provider had refused.
- Resellers: Sellers on a discount inference marketplace read buyers’ prompts via normal request logs.
- Privacy flags: Requests tagged Zero Data Retention (ZDR) were still stored, crawled and in one case shared in a work chat in China.
- For buyers: Every extra hop is one more party holding your full prompt.
How the trap worked
Each prompt carried credentials that existed nowhere else, so any login, DNS lookup or page visit could be traced to one specific request, model, router and privacy setting. A tracing trick using deliberately malformed requests, whose error messages get wrapped by each intermediary, also exposed hops that routers never disclosed.
Three ways prompts escaped
Break-ins. Attackers tested stolen keys, sometimes within hours, searched S3 and AWS secret stores and tried to enable Bedrock models; the decoys drew tens of thousands of AWS API calls. One actor, apparently inside two providers at once, later brought in the LLM-driven pentest framework ARTEX, whose user agents artex-agent and artex-judge-r121 Dragos recommends adding to detections. On retention settings, Moore is blunt: “A privacy flag doesn’t help when the attacker is already inside.”
Reseller logging. No exploit was needed. A seller undercuts the market on a popular model, turns on logging and waits. One person in France tried root SSH logins with passwords lifted from prompts. The risk follows the seller who wins your request, not the whole marketplace.
Ignored settings. A crawler run by a Chinese AI lab visited hidden links from over 50% of the prompts sent to its models and from about two in three requests flagged ZDR or no data collection. US hosts that resell the lab’s models triggered the same crawler.
What it means for teams buying inference
This is the strongest public evidence so far that AI inference behaves like any other data supply chain: whoever touches the payload can keep it. When Google lengthened Gemini chat retention, the debate was about one first-party policy. Dragos shows that a written policy may not survive the second or third hop. And leaks like the DeepAudit flaw that exposed LLM API keys weigh more when agents paste whatever they can read into a prompt.
- Map your route. Ask your router which upstream hosts it may use, pin them yourself and test that a privacy setting really changes the path.
- Skip discount marketplaces for anything sensitive. A price far below list usually means someone is paying another way.
- Keep secrets out of model context. Give agents scoped, short-lived tokens, and prefer designs such as Anthropic’s linked billing, where apps never hold a raw API key.
- Rotate and watch. Rotate any key that has ever appeared in a prompt and plant a canary of your own. Thinkst offers free Canarytokens.
- Verify ZDR. Treat it as a contract term to audit.
i10x Perspective
Cheap inference has a hidden price, and Dragos has now put numbers behind it. Expect enterprise buyers to demand hop-by-hop disclosure from routers, much as they demand SOC 2 reports from cloud vendors. Follow new cases in our AI Security hub.
AI tools assisted with research and drafting; selection, fact-checking and final editing were done by the i10X editorial team.



