Key Takeaways

  • A canary-token experiment found evidence of prompt theft, compromised hosting providers, and ignored privacy settings.
  • OpenRouter removed a Zero Data Retention label from MiniMax endpoints after reports that planted URLs had been accessed.
  • Enterprises may need to treat every intermediary in an AI inference chain as a separate security, privacy, and compliance risk.

Nearly 2 million prompts carrying unique, traceable secrets have exposed a troubling weakness in the AI services market: data submitted to one application may pass through a chain of routers, resellers, hosting companies, and model providers, creating several opportunities for collection or theft.

The chief research officer at industrial cybersecurity company Dragos designed the experiment to trace those movements. According to the Dragos research, a different digital canary was placed in each prompt sent to an AI provider. These included functioning credentials associated with empty Amazon Web Services accounts, SSH passwords, unique URLs, DNS names, and references to private git repositories.

Because every secret appeared in precisely one request, subsequent activity could be linked to a particular route through the inference market. The researcher controlled the receiving infrastructure, making it possible to see when a planted credential was tested or a secret URL was visited.

The results went beyond routine logging. The experiment revealed evidence that attackers had compromised at least three AI providers and were copying prompts. Multiple resellers also appeared to log requests and pass their contents to parties that later used the credentials. Some providers allegedly disregarded privacy indicators such as Zero Data Retention, or ZDR, and "no data collection."

In one breach, attackers reused a couple hundred AWS keys. Every request submitted during several short windows was stolen, including some requests rejected before a model ran. That detail matters. It suggests the copying happened near the provider's front door rather than inside the model itself.

The bait credentials were subsequently used tens of thousands of times in AWS API calls. Those responsible examined S3 buckets, secrets managers, and parameter stores, tried to activate hosted models in AWS Bedrock, and added IAM users in an apparent attempt to preserve access. Credentials sent to two compromised providers were used almost immediately from the same rented server, indicating that the same adversary was likely involved.

Encryption in transit does not resolve this particular exposure. Prompts are decrypted so routers and inference providers can process them. An application might send a request to Vercel, which could pass it to OpenRouter, which could then select a model and hosting environment. Each participant may see plaintext and may retain logs under different policies.

That structure complicates conventional vendor assessments. Which company is actually processing the sensitive data? For many enterprises, the answer may change from one request to another if routing is based on availability, model suitability, or price.

The Ars Technica report also described resellers using low prices to attract traffic before enabling request logging. The researcher isolated reseller behavior by sending prompts to the same model through different sellers, then observing whether credentials were reused only along one route.

MiniMax has become a specific focus following the disclosure. OpenRouter told customers that potential exposure involved prompts sent to MiniMax, including requests routed through AtlasCloud, Novita, and StreamLake. As also reported by Sulat.com, OpenRouter removed the Zero Data Retention label from MiniMax-model endpoints after planted URLs were accessed.

The research found URL fetching days after prompts were submitted, including requests carrying ZDR or no-collection settings. One US hosting provider reportedly crawled a secret URL more than two months after a single inference request. References to private repositories also triggered later cloning attempts in some cases.

For enterprise buyers, a ZDR label should therefore be treated as a contractual and technical claim requiring verification, not as a security control on its own. Organizations can reduce exposure by preventing credentials, customer records, proprietary code, and regulated information from entering prompts. Data-loss prevention checks at the application layer can help, as can short-lived credentials and tightly restricted AWS permissions.

The researcher identified on-premises model hosting as the strongest isolation option, although its cost may put it beyond many organizations. A more accessible approach is to contract directly with a first-tier provider, minimize routing layers, disable default provider pools, and explicitly select upstream processors. That still leaves risk, but it gives security and procurement teams a shorter chain to audit, clearer contractual accountability, and fewer places where a supposedly private prompt can quietly become someone else's data.