OpenRouter logo with headline asking who else read your prompts, about inference provider prompt leaks found by Dragos canary tokens

Inference Provider Prompt Leaks: OpenRouter Drops Three Providers After Dragos Canary Test

Short answer: Inference provider prompt leaks are now documented, not theoretical. Security firm Dragos seeded about two million AI prompts with unique fake credentials and secret URLs, then caught attackers inside at least three hosting providers, marketplace resellers reading customer prompts, and providers that ignored Zero Data Retention (ZDR) flags. OpenRouter, which says it was not breached itself, has off-boarded three providers and pulled 125 endpoints from ZDR eligibility. The caveat: most affected providers are still not named publicly, so treat any secret that ever went into a prompt as exposed.

Key points

  • Dragos sent roughly two million prompts carrying canary AWS keys, hostnames and URLs to widely used AI services over five months and watched where they surfaced.
  • Three providers were compromised. In the largest breach, a couple hundred seeded AWS keys were stolen, and in two breaches some stolen requests were flagged ZDR.
  • Stolen keys were tested within hours, then used to list S3 buckets, search Secrets Manager and Parameter Store, enable Bedrock models and add IAM users, adding up to tens of thousands of AWS API calls.
  • Resellers on an inference marketplace could read prompts simply by turning on request logging. One router quietly forwarded some traffic to that marketplace.
  • OpenRouter removed three providers, removed 125 endpoints from ZDR eligibility and notified affected customers. Together AI, Phala and Novita have published incident notices.

What Dragos found

The research, led by Dragos Chief Research Officer HD Moore, used canary tokens: secrets that exist in exactly one prompt, so any use of them proves that prompt was read by someone. Every request carried an AWS key, a DNS name and a URL that had never appeared anywhere else. Dragos says tokens started firing within days, and the earliest ones are still firing.

The findings fall into three groups:

  • Provider breaches. Attackers were copying requests at at least three hosting providers, none of which knew until Dragos told them. Some stolen keys came from requests the provider rejected before a model ran, which points to theft at the front door, not deep in the inference stack. A ZDR flag does not help when the attacker is already inside.
  • Marketplace sellers. On one inference marketplace, sellers who listed a model at the lowest price could log every prompt routed to them. Dragos watched individuals on home connections open secret URLs and try seeded passwords over SSH by hand. Prompts for the same model sent through a different seller were never touched.
  • Ignored privacy settings. One Chinese AI lab’s systems fetched secret URLs from more than half of the prompts sent to its models, including about two thirds of requests marked “no data collection” or ZDR. Several U.S. hosts that resell that lab’s models triggered the same scraper, so picking a U.S. provider did not keep data in the U.S. In one case, a secret link showed up in a Feishu company chat 24 days after a ZDR-flagged request.

Dragos also built a kind of traceroute for inference and mapped about two thousand routes. Several hundred passed through a system it could not identify, and one marketplace seller was adding about a thousand hidden tokens to every prompt. The attacker behind two of the breaches later ran ARTEX, an LLM-driven penetration testing framework; Dragos lists the user agents artex-agent and artex-judge-r121 as worth adding to CloudTrail detections.

What OpenRouter and the providers changed

OpenRouter says the findings show no breach or unauthorized access of OpenRouter itself. It off-boarded three providers, removed 125 endpoints from ZDR eligibility (the flagged providers plus a precautionary set), and notified every potentially affected customer, including customers of providers that have not posted publicly. It is now adapting Dragos’s testing method to run continuously, checking network paths for undisclosed forwarding, and requiring new assurances before any endpoint gets ZDR status back.

Three providers have published notices so far:

  • Together AI: prompt samples from MiniMax M3, GLM-5.2 and Kimi K2.6 requests were stored in a private Hugging Face repository for testing. An AI coding agent committed a credential for that repository to Together’s public GitHub repo, and an outside party used it to download the samples and test credentials found inside. Together now makes ZDR the default for new accounts and checks every pull request for secrets. If you use agents in your own repos, this is the same failure we covered in Claude Code saving passwords to GitHub.
  • Phala: an attacker took control of an under-hardened inference server. Anyone whose DeepSeek V4 Flash 0731 requests were routed to Phala by OpenRouter during the window in Phala’s report should treat request contents as exposed and rotate any secrets in them.
  • Novita: a routing configuration issue sent some MiniMax requests to MiniMax’s official API instead of Novita’s own hardware for about six months, including requests that arrived through OpenRouter. Novita suspended the affected models.

What the inference provider prompt leaks mean for builders

If you send code, logs or agent context to a hosted model, you are trusting every hop on the route, not just the brand on your invoice. Routers save money by sending each request to the cheapest healthy provider, and Dragos’s point is blunt: if you do not choose where your prompts go, the lowest price chooses for you. Coding agents make this worse because they read .env files, shell history and config, then send whatever they read. That is why the LiteLLM supply chain attack and this story land on the same risk: your API layer now sits next to your credentials.

What to do this week:

  1. Check your OpenRouter email and activity. Look for requests served by Phala (DeepSeek V4 Flash 0731), Novita (MiniMax models) or Together (MiniMax M3, GLM-5.2, Kimi K2.6) in your logs.
  2. Rotate anything that ever sat in a prompt. AWS keys, database URLs, GitHub tokens and passwords pasted into chats or read by agents count. Check CloudTrail for unfamiliar Bedrock enablement or new IAM users.
  3. Pin providers instead of accepting the default pool. In OpenRouter’s provider object, set "zdr": true, "data_collection": "deny", an explicit "only" list of first-party providers you trust, and "allow_fallbacks": false. Remember Dragos’s warning that a privacy flag binds only the first hop.
  4. Seed your own canary. Put a free Thinkst Canarytoken AWS key into the context your agents already send. If it ever fires, you have an incident and a clear trail.
  5. Scrub secrets before prompts leave. Use provider-side secret filtering where it exists, and a local redaction step for anything shaped like a key.

For the most sensitive code, a local model removes the third party entirely. Dragos notes NVIDIA’s DGX Spark now costs $6,950, and small open models such as Inkling 2 Mini, which runs on a 32GB laptop, make that more realistic than it was a year ago.

Our take: discount inference marketplaces were always a data trade, and now there is proof. Going direct to a first-tier provider, or pinning one, is the cheapest security upgrade most teams can make. App builders who store user data have a parallel lesson from the Lovable data exposure: defaults decide who can read your data.

What to watch

  • Whether the unnamed providers, including the one Dragos says was breached first and has not disclosed it, publish incident reports.
  • Which endpoints regain ZDR status on OpenRouter, and what proof OpenRouter asks for.
  • Whether other routers and marketplaces adopt canary testing and disclose every hop on a route.
  • Phala’s full root cause analysis, which it says will follow once remediation is verified.

Sources

Leave a Comment

Your email address will not be published. Required fields are marked *