The short version
If your app sends prompts through an AI router, treat any secret that ever appeared in a prompt as possibly exposed and rotate it. Dragos found attackers copying requests at three inference providers, sometimes within the first hop, and the Zero Data Retention flag did not stop them. Teams building on GenAI integrations should now treat the router and every provider behind it as a data processor with its own breach risk.
What did the Dragos canary study find?
The study found three separate problems in the AI inference supply chain: outright breaches at hosting providers, resellers reading prompts through ordinary logging, and providers that ignored privacy flags. HD Moore, Dragos’s chief research officer, published the results on November 19, 2026.
The method was simple. His team created millions of unique secrets that no legitimate system would ever use, including working AWS keys tied to empty accounts, SSH passwords, private repository references and one-off URLs. Each secret went into exactly one prompt sent to one provider. If a secret was later used, the team knew which hop had read that prompt.
| Finding | What the canaries showed |
|---|---|
| Provider breaches | At least three hosting providers compromised; the largest case leaked a couple hundred AWS keys sent in short windows in late October |
| Reseller theft | Sellers on an inference marketplace listed models cheaply, turned on request logging and reused the secrets they collected |
| Ignored privacy flags | A Chinese AI lab fetched secret links from more than half of prompts sent to its models, including about two-thirds of ZDR or no-collection requests |
| Long tails | One US provider crawled a secret link more than two months after a single request; planted secrets were still being used as of this week |
How did attackers use the stolen prompts?
Attackers used the stolen prompts to go straight for cloud credentials. In the largest breach every request Dragos sent during the affected windows was stolen, including requests the provider rejected before any model ran. That points to an attacker copying traffic at the provider’s front door rather than deep inside the inference stack.
- Test. Stolen AWS keys were checked for validity, in some cases within hours of being planted.
- Search. The attacker listed S3 buckets and searched Secrets Manager and Parameter Store for more credentials.
- Monetize. They tried to enable hosted models in AWS Bedrock and create Bedrock API keys.
- Persist. They added their own IAM users to keep access.
In total the bait keys were used tens of thousands of times in AWS API calls. Secrets from two of the breached providers were used together from the same rented server, so one adversary probably hit both.
On November 1 the same attacker also ran ARTEX, an AI-driven penetration testing framework, to confirm stolen keys were still live; Dragos suggests adding its user agents, artex-agent and artex-judge-r121, to detections.
“A privacy flag doesn’t help when the attacker is already inside,” Moore wrote.
What have OpenRouter and the providers said?
OpenRouter, which routes customer requests to a pool of third-party model hosts, says it removed three providers, took 125 endpoints out of Zero Data Retention eligibility and notified potentially affected customers. It says its own systems were not breached. Endpoints that lost ZDR status will need new assurances and validation before it returns, and OpenRouter is adapting the canary method to run checks continuously.
Dragos has not publicly named the three breached providers. Separately, several providers on the platform have published their own notices:
- Phala says an attacker controlled one of its inference servers and could view requests while they were processed. Requests to DeepSeek V4 Flash 0731 routed through OpenRouter between October 21, 00:00 UTC and November 8, 05:35 UTC should be treated as exposed.
- Together AI says an unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential to a public GitHub repository. The samples came from MiniMax M3, GLM-5.2 and Kimi K2.6 requests made in June and July. Zero Data Retention is now the default for new Together AI accounts.
- Novita says a routing configuration issue sent some MiniMax requests on to MiniMax’s own API. OpenRouter has removed the ZDR label from MiniMax-model endpoints, which researchers traced through Novita, AtlasCloud and StreamLake.
What it means for US & EU software teams
Every router and provider in an AI call path should now be treated as a data processor that can read full prompts. Most AI features in production now carry far more than a user’s question. Coding agents send source files and config, support bots send customer records, and internal copilots send whatever an employee pastes. When that traffic goes through a router, each hop decrypts the full prompt, and the provider that answers may be one you never chose.
For engineering and security leads, three things change.
- Every hop is a subprocessor. Under GDPR Article 28 you are expected to know who processes personal data on your behalf. A router that silently forwards traffic to a marketplace seller or an overseas lab breaks that chain. Data residency promises made to EU customers don’t survive a default provider pool.
- ZDR is a contract term, not a control. It says what an honest provider will store. It won’t help against a breached front end or a reseller that reads logs. Security reviews for SOC 2 and vendor questionnaires should stop accepting a ZDR checkbox as evidence.
- Prompts are now a credential leak path. The attacker here didn’t care about model output. They wanted the AWS keys that developers and agents pasted into prompts. Secret scanning has to cover what leaves your network through AI calls, not just what lands in git.
What should teams using AI routers do this week?
- Rotate. Rotate any key, token or password that may have appeared in a prompt sent through OpenRouter or another router since July, starting with the Phala and Together AI windows above.
- Check AWS. Review CloudTrail for unexpected Bedrock model access requests, new Bedrock API keys, new IAM users and bursts of S3, Secrets Manager or Parameter Store reads.
- Pin providers. Route sensitive workloads to named upstream providers you have a contract with instead of the cheapest available pool. Moore’s advice is to self-host where you can, or buy directly from a first-tier provider.
- Strip secrets before the call. Add a redaction step in your AI gateway or agent tooling that removes credentials, tokens and connection strings from prompts and tool outputs.
- Plant your own canaries. Seed a few unique, monitored credentials in test prompts for each provider you use. It’s cheap, and it tells you exactly which hop leaks.
- Update your records. Add AI routers and the providers behind them to your data processing inventory and vendor risk reviews.
Frequently asked questions
What did the Dragos canary study find?
Dragos sent about two million prompts containing unique AWS keys, passwords and URLs to AI services between July and November 2026. The planted secrets showed that at least three inference providers had been breached, that some marketplace resellers logged and reused customer prompts, and that some providers ignored Zero Data Retention and no-data-collection flags.
Which providers did OpenRouter remove?
OpenRouter said it removed three providers from its platform and took 125 endpoints out of Zero Data Retention eligibility, covering the providers the researchers flagged plus others cut as a precaution. Dragos has not publicly named the three breached providers. OpenRouter says its own systems were not breached.
Does Zero Data Retention protect my prompts?
Not on its own. Zero Data Retention is a policy promise about storage. In two of the three breaches some stolen requests carried the ZDR flag, because the attacker copied traffic as it arrived. A Chinese AI lab also fetched secret links from about two-thirds of the ZDR or no-collection requests sent to its models.
Which requests should be treated as exposed?
Phala says requests to DeepSeek V4 Flash 0731 that reached it through OpenRouter between October 21, 2026 00:00 UTC and November 8, 2026 05:35 UTC should be treated as exposed. Together AI says stored prompt samples from MiniMax M3, GLM-5.2 and Kimi K2.6 requests made in June and July were downloaded. Any secret in those prompts should be rotated.
What should teams using AI routers do now?
Rotate every credential that may have appeared in a prompt, check AWS CloudTrail for unexpected Bedrock activity and new IAM users, strip secrets from prompts before they leave your network, pin specific upstream providers instead of default pools, and review the routers and resellers in your data processing records.
Sources
Dragos — Three breaches, endless leaks: Your AI supply chain needs canaries (primary source, November 19, 2026)
Ars Technica — Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced. (November 20, 2026, updated November 21)
The Hacker News — OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms (November 22, 2026)