AI prompts leaked: OpenRouter drops 3 providers after Dragos study

· Cortex Solutions

Glass data conduit cracked between two servers, golden keys leaking out beside a glowing canary

Dragos planted trackable secrets in nearly two million AI prompts and watched them get used. OpenRouter then dropped three providers. What businesses should do.

When an employee types a question into an AI tool, the words may not go to a single company. Sometimes they pass through a chain of intermediaries and hosts before reaching the model. This week a large field study tested what happens inside that chain, and a well-known routing platform took action after it.

What did Dragos announce?

Cybersecurity company Dragos published a report on November 19, 2026, written by its Chief Research Officer, HD Moore. According to the report, between July and November 2026 the team sent nearly two million prompts to AI providers, each carrying unique, trackable secrets known as "canary tokens": real AWS keys tied to empty accounts, links, domain names and login details.

The idea is simple: the secret exists in one place only, so if anyone uses it, you know that place has been exposed.

Main findings

  • Data from at least three providers was exposed. The report says none of them knew about the breach until the team notified them. It does not name them.
  • In the largest case about two hundred AWS keys were used, each sent in a single request to a single provider during short windows in late October. Keys were sometimes tested within a few hours of being sent.
  • The bait keys were used for tens of thousands of AWS API calls.
  • Sellers on marketplaces that resell model access were caught using customers' secrets. According to the report they needed no vulnerability: they listed the model at the lowest price, enabled request logging and waited.
  • Systems of an unnamed Chinese lab fetched the secret links from roughly two-thirds of the requests sent to its models that were marked "no data collection" or Zero Data Retention (ZDR).
  • The team traced about two thousand request routes; several hundred of them included a system it could not identify.

What did OpenRouter do?

OpenRouter is a platform that routes user requests to multiple providers. In a statement dated November 22, 2026, it said it had:

  • Off-boarded three providers based on the results of its investigations.
  • Removed 125 endpoints from Zero Data Retention eligibility, some flagged by researchers and others removed as a precaution.
  • Notified all potentially affected customers.

It stressed that the findings did not show any breach of, or unauthorized access to, OpenRouter itself. It did not name the three providers. Its statement links to security announcements from Novita, Phala and Together, but it does not say these are the off-boarded companies, and that should not be assumed.

Together AI's disclosure

On November 21, 2026, Vipul Ved Prakash, CEO of Together AI, published a disclosure about an incident involving samples of prompts sent to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, primarily on June 24 and July 4, 8 and 14, 2026. The cause, according to the disclosure, was that an AI coding agent, during development of a public repository, committed a credential giving access to a private repository holding the samples.

The company said five of the credentials that were tested were canary credentials from the Dragos experiment, that it had found no evidence of further leakage, and that its review of the affected content is ongoing.

Before you draw conclusions

The report appears on the blog of a security company with products to sell, and came without raw data or a peer-reviewed paper. It acknowledges that a canary tells you a prompt leaked but not where in the chain. It also says explicitly that none of this means you should stop using AI services.

What does Dragos recommend?

  • Host the model yourself if you can afford it.
  • Or go directly to a first-tier provider, knowing this does not remove the risk.
  • Use routers with care: choose the providers yourself, verify that privacy settings actually change the routing, and stay away from reseller marketplaces.
  • Plant canary tokens in your agents' environment and treat any use of them as an incident.
  • Enable secret filtering at the provider, although it is not a complete solution.

What does this mean for Saudi businesses?

This is Cortex Solutions' reading and does not come from the sources:

  • Know your data's path: ask every vendor who hosts the model, whether requests pass through an intermediary, and where logs are kept.
  • Do not put passwords, keys or customers' personal data in prompts, and write that into a clear policy for your staff.
  • If a password or key passed through an intermediary in a prompt, replace it at its source and review access logs.
  • Do not settle for "we don't keep your data"; ask for it in the contract with a right to verify, and review your obligations under the Personal Data Protection Law with a specialist.
  • Give AI agents the least privilege that does the job, with short-lived keys.

The cheapest price per request is no saving if it costs you one of your company's secrets.

If you would like help building AI automation for your business, this is the service page: https://cortex.sa/services/ai-automation

Sources

  • Dragos, November 19, 2026: https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
  • OpenRouter, November 22, 2026: https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/
  • Together AI, November 21, 2026: https://www.together.ai/blog/security-event-prompt-leakage-for-requests-in-june-and-july-of-2026

Read next