Rotate OpenRouter Secrets After Dragos Finds Prompt Leaks

Dragos seeded about two million canary prompts and found three breached inference providers, logging resellers and ZDR-flagged prompts opened days later. OpenRouter has pulled 125 endpoints from ZDR eligibility.

By Rajesh Beri·November 22, 2026·10 min read
Share:
A rack of inference servers in a dim data hall with a single paper luggage-style tag tied to one network cable, the cable running past the rack into a second, unmarked cabinet.

Illustration generated using AI

If your agents or developers send open-weight model traffic through a router, assume that any credential that ever sat in a prompt has been read by someone you never contracted with, and rotate it. Dragos seeded about two million prompts with unique, trackable secrets between July and November 2026 and found that three inference providers had been breached, that marketplace resellers kept prompts using ordinary request logs, and that prompts marked Zero Data Retention were still copied and opened days later. On November 22 OpenRouter off-boarded three providers and pulled 125 endpoints from ZDR eligibility. A ZDR toggle on a router records what each provider says it does with your data, and the Dragos canaries measured what some of them actually did.

Zero Data Retention (ZDR) is a provider's commitment not to store prompts or responses after processing them. A canary token is a fake credential or URL that exists nowhere else, so any use of it proves that someone read the text it was hidden in.

What Dragos Found in Two Million Canary Prompts

Three providers lost customer data to attackers, and none of them knew until Dragos told them. HD Moore's write-up, published November 19, describes "approximately two million prompts containing unique URLs and canary token credentials" sent over five months, each carrying an AWS key, a DNS name and a URL never used anywhere else. The AWS keys led to empty accounts with full logging, so every attempt to use one left a record.

The findings fall into three groups, and they need different fixes.

The first group is breaches at hosting providers. According to Dragos, "the earliest incident dates back to the middle of July, and none of the three providers knew about their breach until we notified them." In the largest one, "a couple hundred" AWS keys, each sent in a single request, were stolen in a few windows in late October. Some came from requests the provider rejected before any model ran, which puts the copying at the provider's front door rather than inside the model server. In the other two breaches, some stolen requests were marked ZDR. The attackers tested the keys, sometimes within hours, listed S3 buckets, searched Secrets Manager and Parameter Store, tried to enable Bedrock models, create Bedrock API keys and add their own IAM users. Dragos counted "tens of thousands" of AWS API calls against its bait keys, and one key seeded on October 21 was still in use three weeks later.

The second group is marketplace resellers, and no exploit was needed. Sellers, Dragos writes, "only had to list a model at the lowest price, enable request logging, and wait." One attacker on a home connection in France tried an SSH root login with a password taken from a prompt, four times in two minutes. Dragos also found a seller that adds about a thousand hidden tokens to every prompt, and a router that forwarded some traffic to a marketplace without disclosing that hop.

The third group is privacy settings being ignored at the lab end. A Chinese AI lab's systems fetched the secret URLs in more than half of the prompts sent to its models from July to October, including about two thirds of requests marked "no data collection" or ZDR, usually two or three days later. U.S. hosts reselling that lab's models triggered the same scraper. Dragos does not name the lab or the three breached providers.

What the Provider Notices Say

Four providers have published notices, and each describes a different failure. OpenRouter links all four from its announcement but does not say which three it off-boarded, and it says "not all affected providers have posted publicly."

Provider What happened, per its own notice Window What it tells customers
Together AI An AI coding agent committed a credential for a private Hugging Face repo of stored prompt samples to a public GitHub repo; an outside party downloaded the samples and tested credentials found in them Mainly June 24, July 4, 8 and 14, 2026 (MiniMax M3, GLM-5.2, Kimi K2.6) Enable ZDR in account settings; ZDR becomes the default for new accounts
Phala An attacker chained a legacy internet-facing service with management interfaces that trusted loopback access; the model ran on a new GPU server without full production hardening October 21, 00:00 UTC to November 8, 05:35 UTC, a precautionary window; confirmed misuse covers canaries from October 21 and 22 (DeepSeek V4 Flash 0731 via OpenRouter) Rotate all API keys, tokens and passwords sent in request contents
Novita A routing configuration sent some MiniMax requests to MiniMax's official API instead of Novita's own infrastructure May 28 to November 11, 2026, direct and via OpenRouter Revoke or rotate any secrets included in requests
Atlas Cloud A third party opened private links placed in test prompts sent to MiniMax's official API; Atlas says the access did not happen on its infrastructure Q3 and early Q4 2026 Treat any secret sent to MiniMax M3, M2.7 or M3.1 Flash as exposed

Together's notice says five of the credentials the attacker tested were Dragos canaries. Phala credits an unnamed third-party researcher who sent evidence on November 8, and says stolen credentials belonging to one tenant were misused. The Novita and Atlas notices both trace the exposure to requests that reached MiniMax's own API, and both name MiniMax. Dragos does not name the breached providers, so you cannot match each notice to a specific breach in its write-up, and you should not try to.

Together's root cause is the one most engineering teams can reproduce at home: an AI coding agent committed a secret to a public repository. Together says its automated scanning caught the commit and reverted it, but an outside party had already obtained the credential.


Why a ZDR Flag Only Binds the First Hop

A ZDR setting on a router tells the router which endpoints to pick. It cannot tell you what those endpoints, or whoever they forward to, actually do with your text. OpenRouter's ZDR documentation is candid about this. It says the company "works with providers to understand each of their data policies" and in some cases creates special agreements, and that for private endpoints a customer declares as ZDR it "does not independently verify the upstream's contractual behavior." The same page says ZDR "does not apply to plugins and tools you choose to enable, such as web search."

Dragos puts it more bluntly: "A privacy flag doesn't help when the attacker is already inside." It adds that AI agents will send whatever they can read, credentials included. Novita's notice shows the mechanism. A routing path meant for specific customers applied more broadly, and requests that you sent to Novita, directly or through OpenRouter, ended up at MiniMax for more than five months.

The strongest argument for keeping the router is that none of this is unique to it. A direct contract with one provider still leaves you exposed to that provider's breach and its data-sharing deals, which Dragos concedes. Routers also give you failover, one bill, and price competition among hosts of the same open-weight model. Those benefits are real, and the same price competition is what routed prompts to sellers with request logging switched on. Dragos's point is that if you don't choose your providers, the lowest bid chooses them for you.

This follows the Fireworks AI breach notice earlier in November, and the October wave of AI developer platform breaches ran on the same stolen-key economics. In each case the attacker went after keys that grant cloud access, and that is the pattern your detection should look for.

What OpenRouter Changed on November 22

OpenRouter removed three providers and 125 endpoints, and says ZDR eligibility stays off until it gets "appropriate assurances and validation." The post says the 125 "included providers flagged by researchers, plus additional endpoints that we removed as a precautionary measure," and that the findings "did not show any breach or unauthorized access of OpenRouter." It has notified potentially affected customers, is adapting the researchers' testing to run on an ongoing basis, and is checking network paths for requests forwarded to third parties.

Atlas Cloud went further on its side and took every one of its models off OpenRouter until each passes testing, according to its notice. Atlas also says Moore now tests its models four times a day.

Two caveats before you lean on these changes. OpenRouter has not named the providers it removed, so you cannot reconstruct your exposure from the announcement alone. And checking forwarding paths is still described as work in progress. If OpenRouter's ZDR filter is your data control (our Stripe and OpenRouter analysis covers why a toggle is not a contract), the list behind it has just changed, and you should compare it against the live ZDR endpoint list rather than last quarter's.

What to Do About Prompts You Have Already Sent

Start with the secrets, because those are being used now. Dragos says stolen keys and secret URLs "continue to be accessed, even months after the initial incident."

This Week:

  1. Pull your router and provider usage for May through November 2026 and list every model and provider your traffic reached. Match it against the windows in the table above: Novita's MiniMax traffic from May 28, Together's June and July dates, Phala's DeepSeek V4 Flash 0731 window, and Atlas Cloud's MiniMax models.
  2. Rotate every API key, token, password and signed URL that could have appeared in a prompt, an agent's context or a tool result in that period. Phala, Novita and Atlas all give this instruction. Cloud keys go first; the Dragos attackers went straight for S3, Secrets Manager, Parameter Store and Bedrock.
  3. Search CloudTrail for the behaviour Dragos describes: new IAM users, new Bedrock API keys and model-access requests you did not make. If you run Amazon Bedrock, check model enablement in regions you do not use.
  4. In OpenRouter, set ZDR at the account level and add "provider": { "zdr": true } to requests that carry sensitive data. The docs say the per-request flag combines with account settings and cannot switch enforcement off.

This Month:

  1. Write a provider allowlist by name. Most routers let you pin or exclude upstream hosts; use that instead of "cheapest available." Leave reseller marketplaces off the list, as Dragos recommends.
  2. Seed your own canaries. Put a fake cloud key in agent and pipeline context and treat any use of it as an incident. Dragos suggests a free Thinkst Canarytoken. It is the only control here that tells you about the hops your vendor cannot see.
  3. Add secret scanning to prompts and agent outputs, and to the repositories your coding agents can push to. Dragos says provider-side credential scrubbing can't hurt but is "not a panacea"; Together's incident shows why the repository side matters too.

Before Renewal:

  1. Ask each inference vendor where ZDR is enforced, who verifies it, and whether any request can be forwarded to another company. Our AI vendor security review questions cover the contract wording.
  2. Move workloads that handle regulated or secret-bearing data to a first-tier provider contract, or to hardware you run. Dragos prices a 512 GB Mac Studio from $22,299 and an eight-GPU B300 server at $570,000 or more, so self-hosting fits a narrow set of workloads. Our self-hosting break-even analysis shows where it pays.

The Bottom Line

Routers made open-weight models cheap by letting dozens of hosts compete on price per token, and nobody priced in where the prompt went next. The ZDR terms you negotiated with OpenAI or Anthropic carry a contract and an audit trail, while the toggle on a router is a self-reported policy for every host behind it. Dragos says it is working with about a dozen providers and expects that number to grow.

Rotate the keys that were in your prompts this week, then decide which hosts you will allow.

Continue Reading

Share:

Frequently Asked Questions

What did Dragos find in its AI canary prompt study?

Dragos sent about two million prompts containing unique URLs and canary credentials to AI providers from July to November 2026. It found three providers breached without knowing it, marketplace resellers retaining prompts through request logging, and prompts marked Zero Data Retention that were still copied and opened days later.

Which providers did OpenRouter remove after the Dragos research?

OpenRouter says it off-boarded three providers and removed 125 endpoints from ZDR eligibility on November 22, 2026, but it did not name the three. It links incident notices from Atlas Cloud, Novita, Phala and Together, and says not every affected provider has posted publicly.

Does a ZDR setting on OpenRouter guarantee my prompts are not stored?

No. OpenRouter routes to endpoints based on each provider's stated data policy. Its docs say it does not independently verify upstream behaviour for private endpoints a customer declares as ZDR, and ZDR does not cover plugins such as web search. Dragos found ZDR-marked requests that were retained and read.

What should I do if my prompts went through an affected provider?

Rotate any API key, token, password or private URL that appeared in a prompt or agent context, starting with cloud keys. Check CloudTrail for new IAM users, Bedrock API keys and model-access requests you did not make, then pin a named provider allowlist and seed your own canary credentials.

Newsletter

More on Models & LLMs, every Sunday

The week's enterprise AI news, Models & LLMs included, in one Sunday email.

Weekly, free. Unsubscribe any time. Privacy

Latest Articles

View All →