Inter-office memo

To
Everyone sending prompts through OpenRouter, Together AI, Phala, Novita or MiniMax
From
Mo, Chief Executive
Date
November 21, 2026, 3:10 PM CST
Re
OpenRouter’s provider incidents

OpenRouter drops three providers over leaked prompts. Who’s affected, and what to do

On November 22, 2026, OpenRouter said it had off-boarded three unnamed providers and pulled 125 endpoints’ Zero Data Retention label after Dragos’s bait-prompt research. Together AI says an AI coding agent leaked a credential to its prompt samples; Phala, that DeepSeek V4 Flash 0731 requests were exposed; Novita, that MiniMax requests went to MiniMax. If you used them, replace every secret those prompts carried.

What OpenRouter did on November 22

On November 22, 2026, OpenRouter, which passes requests from one API to many providers’ models, published Data Processing Incidents on Inference Providers. “Earlier this week, security researchers reported that certain inference providers had data compromised or otherwise shown evidence that their data policies were not being upheld,” it says, linking Dragos’s research, and “These findings did not show any breach or unauthorized access of OpenRouter.” It has:

  • “Off-boarded three providers from OpenRouter, based on the results of our investigations.”
  • “Removed 125 endpoints from Zero Data Retention (ZDR) eligibility. The removals included providers flagged by researchers, plus additional endpoints that we removed as a precautionary measure.”
  • “Notified all potentially affected customers.”

It is also adapting the researchers’ tests to run on an ongoing basis, checking “network paths to identify instances where requests and responses may be being forwarded to third parties,” and “requiring appropriate assurances and validation before restoring ZDR eligibility to any endpoint removed in this review.”

It doesn’t name the three providers or list the 125 endpoints. It links incident notices from Novita, Phala and Together AI, and adds: “Not all affected providers have posted publicly. We have notified affected users regardless of whether there is a public post.” As the security site The Hacker News puts it, “Dragos has not named the three breached providers, OpenRouter has not named the three it removed, and neither has said publicly whether they are the same three.” OpenRouter’s own pages, on November 22: Phala’s and AtlasCloud’s listed no models, Novita’s listed 35 with none from MiniMax, and Together’s and StreamLake’s listed 26 and 18, both including MiniMax M3.1 Flash.

What Together, Phala and Novita disclosed

Provider What happened Requests affected What it says to do
Together AI An AI coding agent committed a credential to a public GitHub repository, and someone used it to download stored prompt samples Samples of MiniMax M3, GLM-5.2 and Kimi K2.6 requests to its serverless API, mainly from June 24, July 4, July 8 and July 14, 2026 Turn on Zero Data Retention in account settings; questions to security@together.ai
Phala An attacker took over the server running one model, so requests were visible while they were processed DeepSeek V4 Flash 0731 requests OpenRouter sent it, October 21 at 00:00 UTC to November 8 at 05:35 UTC Rotate every secret in those requests, and check those accounts for misuse
Novita A routing setting sent MiniMax requests to MiniMax’s official API instead of its own servers MiniMax requests, direct or through OpenRouter, May 28 to November 11, 2026 Revoke or rotate any secrets in them; questions to privacy@novita.ai

Together AI. Its notice, posted on November 21 by its CEO, Vipul Ved Prakash, covers “a subset of prompts submitted to Together AI’s serverless inference service in June and July 2026.” Together had “collected prompt samples authorized for that use” for testing and kept them in a private Hugging Face repository. Then, “During development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repository.” Its scanning caught the credential and it reverted the commit, but “During the exposure window, an unauthorized party obtained the credential and used it to download the stored prompt samples.” That party “searched those samples for credentials to other services and tested whether those credentials were valid. Five of the credentials tested were canary credentials associated with a Dragos security experiment.”

Together says it has “not identified evidence of additional prompt leakage beyond this incident,” is still reviewing what the samples contained, and “will contact affected customers directly as needed.” Its fixes: scoped or short-lived credentials, “Checking all pull requests for credentials,” “Making zero data retention (ZDR) the default setting for new accounts,” and “Updating our development processes governing the use of AI coding agents.” The notice gives no dates for the exposure window and doesn’t mention OpenRouter.

Phala. Its preliminary report is dated November 10, nine days before Dragos’s. “On November 5, we detected unauthorized activity on one server,” it says: an attacker had chained “a legacy internet-facing service with management interfaces that relied on loopback access for trust and did not independently authenticate callers,” creating unauthorized confidential virtual machines and taking application credentials. On November 8, “a third-party security researcher provided evidence that request data sent to DeepSeek V4 Flash 0731 had been exposed,” which Phala traced to “a new GPU server. It was serving production traffic without our full production security hardening.” The attacker had taken control of it, “making in-flight request data visible during processing.”

The researcher had put bait credentials in requests to Phala from July 1 to November 7, and “Only canary credentials from requests sent on October 21 and 22 were subsequently misused.” Phala keeps the wider window because the tests “did not provide visibility into every request.” It says its audit of the whole platform found no compromise of other models or products, that later attempts by the attacker failed, and that a full root cause analysis will follow. Dragos wrote on November 19 that one of its three breached hosts “has already published a preliminary security notice”; neither company has said that’s Phala’s.

Novita. Its notice, dated November 22, says “some requests were forwarded to MiniMax’s official API rather than processed on Novita-hosted infrastructure,” because “A routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly.” “If you used MiniMax models through Novita between 2026-05-28 and 2026-11-11, your request contents may have been accessed or retained by a third party. This includes requests made directly through Novita and those routed to Novita through OpenRouter.” It has suspended the affected models. The researcher saw test URLs and hostnames accessed but “No use of the AWS test key markers,” and “The access patterns suggest automated content review may have been involved,” which Novita is checking with MiniMax, along with “any retained customer data and its deletion.” Its window, May 28 to November 11, is wider than the researcher’s July-to-October test.

What OpenRouter’s first email said

Before the announcement, OpenRouter emailed customers under the subject “Requests you sent to MiniMax models may have been read.” It didn’t publish that notice; a customer posted a copy on Hacker News at 10:31 PM CST on November 20, 2026, noting that “There is no public page or blog post,” and others in the thread said they had received it too. The email says:

  • “A security researcher placed unique test URLs in prompts sent to MiniMax models from July to October 2026.”
  • “A third party accessed those URLs 40 to 130 hours after the requests were sent, whether MiniMax served a request directly or another provider forwarded it to MiniMax.”
  • “AtlasCloud, Novita and StreamLake have told us they forwarded OpenRouter requests for MiniMax models to MiniMax.”
  • “With one exception, the same tests on other models at those providers were not accessed, and none of the AWS test keys were used.”
  • “MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22 and that it has since enabled Zero Data Retention. We have asked MiniMax for its own account and have not received one.”
  • “This means a third party may have read the prompts and outputs of these requests.”
  • “While we investigate, we have removed the Zero Data Retention label from these MiniMax-model endpoints.”

The posted copy names that account’s own requests: “Your account sent [a number of] requests to MiniMax models through” the providers it used, between two times in UTC. OpenRouter placed the fault elsewhere: “the issue sits with MiniMax and the providers that forwarded to it, not with OpenRouter.” It didn’t name the researcher, the third party or the one exception, and it didn’t ask customers to do anything.

The research behind it

On November 19, Dragos published a report by HD Moore, its chief research officer. “Over a five-month period from July through November 2026,” his team sent about two million prompts to AI services, each carrying bait: unique URLs, DNS names, passwords, git repository references and canary AWS keys. Then they watched where the bait turned up. Three findings matter here:

  • A Chinese AI lab’s systems followed links in prompts, including ones marked private. “This lab’s systems fetched the secret URLs from more than half of the prompts that we sent to its models between July and October, including roughly two-thirds of the requests marked as ‘no data collection’ or Zero Data Retention.” “The scraper typically showed up two or three days after the request.”
  • U.S. hosts didn’t keep prompts in the U.S. “Several U.S. hosting providers resell this lab’s models, and prompts sent to those providers triggered the exact same scraper – the prompts were flowing back to the lab.” In Dragos’s words, “A customer who picked a U.S. provider specifically to keep their data in the U.S. didn’t get what they paid for.”
  • Three hosting providers were breached. Attackers used the bait AWS keys “for tens of thousands of AWS API calls,” went looking for data in S3, Secrets Manager and Parameter Store, and “tried enabling hosted models in AWS Bedrock, creating Bedrock API keys, and adding their own IAM users to keep access.” “The earliest incident dates back to the middle of July.”

Dragos names neither the lab nor the hosts. OpenRouter’s announcement now links its report as the research behind the review and thanks Dragos for it, and Together says five of the credentials tested were Dragos’s; Phala and Novita credit an unnamed researcher. The match between Dragos’s Chinese lab and MiniMax is in the details: the same July-to-October window, fetches two or three days after a request against 40 to 130 hours, and resellers passing requests to the lab. Coverage such as The CyberSec Guru’s draws that link; neither Dragos nor OpenRouter has stated it.

Who it affects

If you What’s been said What to do
Got an email from OpenRouter The requests it lists may have been read or exposed Treat them as read, and replace every secret in them
Used Together AI’s serverless API in June or July 2026 Samples of MiniMax M3, GLM-5.2 and Kimi K2.6 requests were downloaded; Together will contact customers “as needed” Replace secrets those requests carried, and turn on Zero Data Retention
Sent DeepSeek V4 Flash 0731 requests through OpenRouter, October 21 to November 8 Phala says the ones it served are exposed Check which provider served them, and rotate the secrets in Phala’s
Used MiniMax models at Novita, directly or through OpenRouter, May 28 to November 11 They went to MiniMax’s API and “may have been accessed or retained by a third party” Revoke or rotate the secrets in them
Used MiniMax models elsewhere, or MiniMax’s own API Nothing from MiniMax; the researcher’s test covered requests MiniMax served directly, through OpenRouter Keep secrets out until MiniMax explains
Used other models at AtlasCloud, Novita or StreamLake “With one exception,” the tests weren’t accessed; OpenRouter hasn’t said which Nothing yet
Send any open model’s prompts to a host that resells it Dragos found one lab’s prompts flowing back to it through U.S. resellers Choose providers yourself, as below

OpenRouter’s MiniMax page lists 16 models, among them MiniMax M3, M3.1 Flash and M2.7. On November 22, the providers on MiniMax M3’s page were DeepInfra, CoreWeave, Venice, Parasail, Featherless and MiniMax itself; OpenRouter hasn’t said whether hosts that run the open weights on their own servers were part of the test.

What to do

  1. Read every notice you got. OpenRouter’s email gives a count, the providers and a window in UTC. Phala says “The gateway is notifying affected users,” and Together that it “will contact affected customers directly as needed.” To see which provider served a particular request, OpenRouter’s generation endpoint returns its provider_name, “Name of the provider that served the request,” and provider_responses, “including fallback attempts.”
  2. Replace every secret those prompts could have carried. Whoever took Together’s samples searched them “for credentials to other services” and tested what they found. A coding agent sends the files it reads and the output of the commands it runs, so a key in a .env file it opened counts as much as one you pasted. Revoke each with whoever issued it: deleting a log doesn’t stop a copied key from working.
  3. Treat the rest as disclosed. Private code, customer data or anything else in those requests may have been read, so tell whoever needs to know.
  4. Restrict where requests go. OpenRouter’s privacy settings can “only allow you to route to endpoints that have a Zero Data Retention policy,” which can be enforced separately for Anthropic, OpenAI, Google, SpaceXSI and all other models. Per request, provider routing takes zdr (“Restrict routing to only ZDR (Zero Data Retention) endpoints”), data_collection (“Control whether to use providers that may store data”; set it to "deny"), ignore for providers to skip, only for the ones to allow, and allow_fallbacks to stop it trying backups. On Together itself, “Customers handling sensitive prompt data should enable ZDR in their account settings.”
  5. Check what your coding agents commit. Together’s leak began when “an AI coding agent committed a credential” to a public repository. Read an agent’s changes before it pushes, keep keys in files your repository ignores, and see our guide to keeping them away from agents in the first place.
  6. Plant your own canary. Dragos’s advice: “Add a fake credential to the context that your agents and pipelines already send and treat any use of it as an incident.” It adds that “A free Canarytoken from Thinkst is enough to get started today.”

What a Zero Data Retention label promises

OpenRouter’s definition: “Zero Data Retention (ZDR) means no routine storage of prompts or responses after processing, apart from the CSAM exception described on this page.” The same page gives October 27, 2026 as the date it would begin screening all requests, ZDR traffic included, for child sexual abuse material in flight. It decides which endpoints get the label by working “with providers to understand each of their data policies,” sometimes with “special agreements,” and when it can’t establish a clear policy it assumes “the endpoint both retains and trains on data.” So the label is the provider’s promise, passed on, and OpenRouter now requires assurances and validation before it hands the label back to the 125 endpoints it took it from.

These incidents show three ways a prompt can get out. Novita’s MiniMax requests went to a company that hasn’t explained its handling of them; MiniMax’s account, as OpenRouter relays it, is only that content moderation was on before October 22. Together’s prompts were stored on purpose, as test samples, and stolen with a leaked credential. Phala’s were visible to an attacker on its server “during processing,” which no retention policy covers. Dragos’s summary of what it saw: “Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people.” It isn’t the first such gap this month: Notion said Fireworks AI had given it “a contractual commitment to zero data retention,” then found data retained from the endpoints it used (our memo).

What hasn’t been said

  • Which three providers OpenRouter removed, which 125 endpoints lost the label, and when any get it back.
  • Who read the MiniMax requests, and MiniMax’s own account. OpenRouter has asked for one; we found no public statement from MiniMax.
  • The dates of Together’s exposure: when the credential was committed and reverted, how many prompts and customers the samples covered, and whether accounts with Zero Data Retention on were sampled.
  • Who downloaded Together’s samples, and who attacked Phala.
  • The one exception in OpenRouter’s email: which model, at which provider.
  • Whether customers of MiniMax’s own API outside OpenRouter were affected.
  • Whether Dragos’s three breached hosts are among the providers named here.

Keeping prompts on servers you chose

Dragos’s advice comes down to choosing the route: host it yourself “if you can afford it,” go “directly to a first-tier provider” (“one company, one contract, and one privacy policy to read”), pick a router’s providers yourself and “verify that your privacy settings actually change where your requests go,” and “Stay away from reseller marketplaces – some of that capacity comes from stolen keys.” “None of this means that you should stop using AI services,” it says. “It does mean that where your prompts go is a decision, and if you don’t make it, the lowest price will make it for you.”

Enjoy doesn’t sit on that route. It has no AI model of its own and hands your conversations to the agent on your computer (what AI providers see), so if you ran MiniMax or DeepSeek through OpenRouter in OpenCode in Enjoy, these notices apply to you just the same. To send nothing at all, use a local model, which Enjoy runs through OpenCode. MiniMax-M3’s open weights have “~428B parameters and ~23B activated parameters,” which is server hardware, so smaller local coding models are the practical way for most people. We’ll update this memo as the providers and MiniMax say more.

10.Questions.

Which providers did OpenRouter remove?

It hasn’t said. On November 22, 2026, OpenRouter said it had “off-boarded three providers” and removed 125 endpoints from Zero Data Retention eligibility, and it linked incident notices from Novita, Phala and Together AI without calling them the three. That day its pages for Phala and AtlasCloud listed no models, while Together’s listed 26 and Novita’s 35.

What happened at Together AI?

Together says an AI coding agent working on its public GitHub repository, xorl, committed a credential that opened a private Hugging Face repository of prompt samples from its serverless API. Someone used it to download samples from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, mostly from June 24, July 4, July 8 and July 14, 2026, and tested the credentials in them; five were Dragos’s bait. Together now makes Zero Data Retention the default for new accounts.

Were my DeepSeek requests through OpenRouter exposed?

If OpenRouter sent your DeepSeek V4 Flash 0731 requests to Phala between October 21, 2026 at 00:00 UTC and November 8 at 05:35 UTC, Phala says to “treat request contents as exposed” and rotate any secrets in them. An attacker had taken over the server running that model. Phala says its audit found no compromise of its other models, and OpenRouter’s generation endpoint shows which provider served each request.

Were my prompts to MiniMax models read?

If OpenRouter emailed you about them, assume so. Its email says “a third party may have read the prompts and outputs” of the requests it lists, sent to MiniMax models through MiniMax or a provider that forwarded them there. Novita adds that MiniMax requests it took from May 28 to November 11, 2026, directly or through OpenRouter, went to MiniMax’s official API and “may have been accessed or retained by a third party.”

Is MiniMax safe to use now?

Nobody can say yet. MiniMax hasn’t given OpenRouter its own account, and we found no public statement from it. Novita says the access patterns “suggest automated content review may have been involved” and is checking that with MiniMax. Until MiniMax explains, keep passwords, keys and private code out of prompts to its models.

How do I make OpenRouter use only Zero Data Retention providers?

Turn on Zero Data Retention enforcement in OpenRouter’s privacy settings, or set zdr to true in a request’s provider options. Requests then go only to endpoints labeled Zero Data Retention, “plus Safety Retention endpoints if your settings allow them.” The label is each provider’s promise, and OpenRouter says it will require “appropriate assurances and validation” before giving it back to the endpoints it removed.

Enjoy Agentic Industries · Checked against its sources on