A notification email, two million booby-trapped prompts, and the gap between a privacy label and a privacy boundary.
On November 21, 2026, an email from OpenRouter started circulating on Hacker News under the subject line “Requests you sent to MiniMax models may have been read.” One recipient posted it in full, and Max Woolf, the researcher known online as minimaxir, confirmed he had received the same message independently. The email describes an experiment rather than a breach in the usual sense. I think that is exactly why it deserves a slow read.
What OpenRouter told customers about MiniMax prompts
A security researcher placed unique test URLs inside prompts sent to MiniMax models between July and October 2026. A third party then opened those URLs 40 to 130 hours after the requests were sent. It happened whether MiniMax served the request directly or another provider forwarded it to MiniMax. The URLs existed nowhere else, and nobody legitimate had any reason to visit them. The only explanation is that the prompt content was stored, processed later, and acted on by a person or system outside the relationship between the customer and the provider they thought they were using.
The notification was personalized. One affected developer’s copy listed 321 requests to MiniMax models sent through four providers (MiniMax itself, Novita, StreamLake, and AtlasCloud) between June 6 at 09:16 UTC and October 22 at 06:48 UTC. According to the email, AtlasCloud, Novita, and StreamLake told OpenRouter that they forwarded OpenRouter requests for MiniMax models to MiniMax. Based on what the researcher and those providers reported, OpenRouter said the issue sat with MiniMax and the forwarding providers, not with OpenRouter itself.

The email also relayed MiniMax’s explanation. MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22, and that it has since enabled Zero Data Retention (ZDR). OpenRouter asked MiniMax for its own account and had not received one. While it investigates, OpenRouter removed the ZDR label from the affected MiniMax-model endpoints. As of the Hacker News posting there was no public blog post or incident page, so the email was the disclosure.
Two details in the message shape everything else. First, with one exception, the same tests run against other models at those same providers were never accessed. The exposure followed the MiniMax model path, not the providers’ infrastructure in general. Second, none of the researcher’s AWS test keys were used on that path. This was not credential theft. Prompts were retained, reviewed, and had their embedded links followed days after the inference call had returned its tokens and the connection had closed.
How HD Moore’s two million canary prompts work
The researcher behind the email is HD Moore, the creator of Metasploit and now chief research officer at Dragos. The MiniMax finding is one strand of a much larger study. In a writeup published November 19, covered the next day by Ars Technica, Moore describes sending roughly two million prompts containing unique, trackable secrets through widely used AI services between July and November 2026, then watching where those secrets surfaced. The method is a scaled-up canary token: a secret that exists in exactly one place, so any use of it anywhere proves that place has been exposed.
The engineering is what turns anecdote into evidence. In July the team started with Thinkst Canarytokens, seeding each request with an AWS key, a DNS name, and a URL that had never appeared anywhere else. Tokens began firing within days. That proof of concept worked well enough that they built a dedicated seeding service. For every prompt, a unique seed derives unique hostnames, URLs, and credentials, and the service runs the infrastructure those artifacts point to: DNS, web, git, SSH, databases, mail, and a long tail of older protocols. The AWS keys are real keys attached to empty accounts with full logging.
Every component traces back to its parent seed, so each inbound touch can be attributed to the exact prompt, model, router, and privacy setting that carried it. The team records precise timing and sequencing, including which DNS resolver was used, how long it took to reply, and the full fingerprint of the inbound connection from TCP through TLS SNI to HTTP Host headers and requested URLs. A URL fetch two days after a request is direct evidence of retention.
The reason this works is structural, and Ars Technica states it most cleanly: prompts are encrypted in transit but decrypted at every hop. A router reads plaintext to decide where to send it. A hosting provider reads plaintext to run inference. Requests can also pass from one gateway to another. The study found chains where a request entered through one router and quietly exited through a marketplace. At each stop the operator can log, store, or forward the content under whatever policy actually governs that hop. Moore’s summary is that a privacy flag only binds the first hop.
📬 Stay Ahead of Cyber Threats
Get the latest cybersecurity news, critical vulnerabilities, threat intelligence, tutorials, and exclusive giveaways delivered straight to your inbox. No spam. Unsubscribe anytime.
What the MiniMax path showed: prompts retained and read days later
Dragos anonymizes the companies in its report, but the finding it describes matches OpenRouter’s notification closely, and OpenRouter names the lab. In the Dragos account, this Chinese lab’s systems fetched the secret URLs from more than half of the prompts sent to its models between July and October. That included roughly two thirds of the requests explicitly marked “no data collection” or ZDR. The scraper typically appeared two or three days after the request, which fits the 40 to 130 hour window OpenRouter measured.
The behavior was not limited to requests sent to the lab directly. Several US hosting providers resell the lab’s models, and prompts sent to those providers triggered the same scraper, so the prompts were flowing back to the lab. Dragos concludes that a customer who picked a US provider specifically to keep data in the United States did not get what they paid for. Later, a git client on a different cloud began trying to clone the private repositories referenced in the prompts, again only for prompts that used this lab’s models. That is a second differential signal pointing at the model path rather than the resellers.
The timing says something technical. A fetch that arrives 40 to 130 hours after inference cannot be part of serving the request, because the response was delivered long before. It implies content that was persisted in logs, a moderation queue, or some other store, and then processed asynchronously by a batch scanner, a human reviewer, or a pipeline that renders text and follows embedded links.
Here is what is confirmed and what is not. The canaries prove access to prompt content, and OpenRouter’s email extends the warning to outputs, saying a third party “may have read” both. The public record does not establish who or what performed the fetches, or on whose instructions. MiniMax’s explanation, relayed through StreamLake, is that content moderation was enabled during the window and that it has since enabled ZDR. OpenRouter has no first-party account from MiniMax. Its current provider logging table lists MiniMax, AtlasCloud, and StreamLake alike as retaining prompts “for unknown period” while not training on them, and the ZDR labels remain off.
Next to the rest of the study, the MiniMax case sits at the mild end. No credentials were stolen on this path and no attacker needed an exploit. My inference, which fits both the relayed moderation explanation and the observed behavior, is that retention and safety processing happened at the model’s origin, out of sight of a customer who bought the model from a reseller under a different flag. That remains an inference until MiniMax says otherwise.
Novita, AtlasCloud and StreamLake: the forwarding providers
The three forwarding providers are not interchangeable, and the popular framing of this incident as “US neoclouds shipping your data to China” is only partly right. Novita AI is headquartered in San Francisco. AtlasCloud is, per its own privacy policy, Atlas Cloud AI LLC, a New York-registered company at 1540 Broadway, with its contact page listing headquarters in New York City and offices in Silicon Valley and Calgary. StreamLake is the AI cloud brand of Kuaishou, the Chinese short-video company. It launched in 2022 and now sells model-as-a-service inference, including its KAT-Coder line. Two of the three forwarders are US companies and the third is Chinese. All three did something the router’s endpoint metadata never captured: when a request arrived for a MiniMax model, they passed it to MiniMax.
In at least one case, the architecture was documented in plain sight. AtlasCloud’s “How Atlas Cloud Works” page describes the company as an AI API aggregation platform sitting between your application and multiple model providers, with a diagram showing chat completions routed onward to upstreams that include MiniMax, DeepSeek, Alibaba, and Moonshot. Novita’s trust center says ZDR is the default for all endpoints, including serverless API calls, that inputs and outputs are processed transiently and not retained beyond delivery, and that technical metadata is kept for six months. Its sub-processor FAQ also lists a “Model provider” category that provides model inference for integrated models and processes API inputs and outputs per your requests.
The contractual vocabulary for forwarding existed. What did not exist was any per-request visibility, on OpenRouter or anywhere else, telling a customer that a particular completion, labeled ZDR at the endpoint they selected, would run on another company’s infrastructure under another company’s policy.
MiniMax’s terms and jurisdiction
The jurisdictional picture is more tangled than the slogans allow. MiniMax is headquartered in Shanghai and listed on the Hong Kong Stock Exchange in January 2026, one of the first mainland Chinese generative-AI companies to do so. Its international API platform is operated under terms of service by Nanonoble Pte. Ltd., a Singapore-registered entity, with Singapore governing law and SIAC arbitration.
Those terms are candid about data use. Customers keep ownership of their inputs and generated content, but MiniMax “may use the input and generated content to provide, maintain, develop, and improve our Services,” and may access and process client data for maintenance, troubleshooting, support, and legal compliance. A confidentiality clause also says that using confidential information to improve algorithms or enhance services is not a breach. None of this is hidden, and it matches the usual posture of a lab operating under Chinese generative-AI regulation, which carries its own content and logging obligations.
The terms were never the secret. The problem is that a customer who chose a San Francisco or New York provider under a ZDR label was accepting terms they had never read, from a company they had never chosen.
What OpenRouter’s zero data retention flag actually enforces
To see why the label failed, look at what it is mechanically. OpenRouter routing is controlled by a provider object in the request body. Its ZDR documentation describes enforcement at three levels: account-wide privacy settings (configurable per model group, such as Anthropic, OpenAI, Google, xAI, and everything else), guardrails attached to API keys or organization members, and a per-request flag. The request-level control composes as an OR with the stricter layers, so it can force ZDR on but never off.
json
{
"model": "meta-llama/llama-3.3-70b-instruct",
"messages": [{ "role": "user", "content": "..." }],
"provider": {
"zdr": true,
"data_collection": "deny",
"only": ["deepinfra"],
"allow_fallbacks": false
}
}
When zdr is true, the router restricts the request to endpoints whose policy is Zero Data Retention. Separately, data_collection: "deny" excludes endpoints that store data non-transiently and may train on it. The eligible-endpoint list is published programmatically at /api/v1/endpoints/zdr and updates as policies change. When OpenRouter cannot establish a clear policy for an endpoint, it takes the conservative position and classifies it as retaining and training. That is a useful control surface.
It is also made of endpoint-level policy metadata, assembled from provider agreements and attestations and sometimes negotiated into stricter-than-default contracts. OpenRouter’s documentation on private endpoints says so directly: a ZDR declaration is your attestation about your agreement with the upstream, and the platform does not independently verify the upstream’s contractual behavior. The public provider data carries a similar disclaimer, described in independent analysis as not a definitive source of third-party data policies but the platform’s best knowledge. A filter over best-knowledge metadata enforces contracts. It is not a security boundary, and it models the first hop rather than the whole graph.
Safety Retention and the shifting meaning of “zero”
The industry has renegotiated the definition of “zero” in the weeks around this incident. On October 18, Google began requiring what OpenRouter calls Safety Retention for Google models served through Vertex: prompts flagged by safety classifiers may be retained for up to 90 days for abuse monitoring, without training. On October 26, OpenRouter formalized Safety Retention as a third policy tier between ZDR and Retention-and-Training. Endpoints in that tier may retain prompts and responses from any request, flagged or not, solely to detect, investigate, and prevent abuse. Self-serve accounts allow Safety Retention endpoints by default when ZDR is enforced, while enterprise accounts default to ZDR-only. An update planned for October 27 extends in-flight CSAM screening to all OpenRouter traffic, including ZDR requests, with flagged content retainable for human review, reporting, and legal preservation.
Caching draws another fine line. In-memory prompt caching at the provider counts as compatible with ZDR because nothing is written to persistent storage. OpenRouter’s own response caching does store completions, and account-level ZDR disables it but the per-request flag does not.
None of these choices is unreasonable. Together they show that ZDR is a negotiated bundle of exceptions, and the bundle lives in metadata and contracts, which is the layer a silent forwarder bypasses. If your requirement is location rather than retention, OpenRouter offers a separate and stronger control: in-region routing through eu.openrouter.ai and us.openrouter.ai on Business and Enterprise plans, where prompts and completions are processed entirely within the region. Retention and residency are different questions, and this incident shows what happens when a product treats them as one.
Other findings in the Dragos study
The MiniMax scraping would be a manageable story alone. The Dragos research hit harder because of everything around it.
Provider breaches and stolen AWS canary keys
The study found evidence that at least three hosting providers were compromised by attackers copying customer prompts. The earliest intrusion dates to mid-July, and none of the three providers knew until Dragos notified them. In the largest breach, a couple hundred AWS canary keys were stolen, each seeded in exactly one request within short windows in late October, and every request sent during those windows was taken. Some of the stolen keys came from requests the provider rejected before a model ever ran, which places the theft at the provider’s front door, in the ingress or logging layer, rather than deep in the inference stack. In the other two breaches, some of the stolen requests were marked ZDR. As Moore puts it, a privacy flag does not help when the attacker is already inside.
After the theft, the attackers followed a textbook cloud-exploitation path, and the telemetry is useful to defenders. Keys were validated within hours, then used to enumerate S3 buckets, Secrets Manager, and Parameter Store. The actors tried to enable hosted models in AWS Bedrock, created Bedrock API keys, and added their own IAM users for persistence. The bait keys were used for tens of thousands of AWS API calls in total.
One adversary appears to have compromised two of the three providers at once, since secrets sent to both were used together, almost immediately, from a single rented server. Over three weeks of observation the tooling changed from a script spoofing Chrome on Windows, to Boto3 running inside OrbStack, to the AWS CLI on Kali Linux. On November 1 the attacker added ARTEX, an LLM-driven penetration-testing framework whose user agents, artex-agent and artex-judge-r121, appear in no public indicator list and are worth adding to detections. One key seeded on October 21 was first used the next day and was still in use three weeks later. These tools often announce themselves in CloudTrail user agents, but only a fraction of the traffic does.
Reseller marketplaces and undisclosed hops
Below the breaches sits a layer that needed no exploit at all. Marketplace resellers listed popular models at the lowest price, enabled request logging, and waited. The economics are the vulnerability: price-weighted routing hands your prompt to whoever undercuts the field, and some of that cheap capacity comes from pooled consumer subscriptions, startup credits, or stolen credentials.
Dragos caught individuals on home broadband connections opening secret URLs in desktop browsers a day and a half after inference and guessing usernames at login pages. One attacker on a Mac in France hand-typed root SSH passwords taken from prompts, four attempts in two minutes, then tested AWS keys in the same session. Prompts for the same model sent through a different seller were never touched, which shows the leak followed the seller. One key had been sent through an entirely different router that turned out to be quietly forwarding some traffic to the marketplace, an undisclosed hop.
To map chains like that, the team built something like traceroute for inference: image URLs that fingerprint every fetcher, plus deliberately malformed requests whose errors come back wrapped in three or four nested envelopes, with each router’s wrapping style matched against a signature database. Across roughly two thousand traced routes, several hundred included unidentified systems. One marketplace seller was silently injecting about a thousand hidden tokens into every prompt, and another fetched images through a shared consumer subscription account.
The most human anecdote involves a different Chinese provider. Twenty-four days after a prompt was sent through a US-based provider with the ZDR flag, the secret link appeared in a Feishu chat channel and was opened within two minutes on three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the first eight characters of that link, so whoever shared it was reading a stored copy of the full prompt more than three weeks after the request.
Even large US providers were not clean. One kept crawling a secret URL more than two months after a single inference call to its search endpoint, and another runs a scraper that returns days after each request regardless of which router delivered it.
How to protect your prompts on AI gateways
Treat affected requests as disclosed
If you received the OpenRouter email, assume the contents of the affected requests were read, because the notification says they may have been. Anything sensitive that traveled in those prompts between June and October should be rotated or assessed on that assumption. That covers credentials, connection strings, internal URLs, customer data, and unpublished code. The unused AWS test keys tell you how this particular path behaved. They say nothing reassuring about your own secrets, and a moderation pipeline that follows links is still a reader.
Pin upstreams and enforce ZDR on OpenRouter
Where your prompt goes is a decision you have to make explicitly, because otherwise the lowest price will make it for you. On OpenRouter, that means:
- Pin upstreams with
provider.onlyand disable silent failover withallow_fallbacks: false. - Enforce
zdr: trueanddata_collection: "deny"for sensitive traffic. - Check the live
/api/v1/endpoints/zdrlist instead of relying on a cached mental model of which endpoints qualify. The MiniMax labels changed overnight, and the list is the source of truth. - Record the generation ID and serving provider of every request, so the next time a disclosure lands you can answer “were we affected” with a query instead of a guess.
- Use the in-region domains when processing location is the actual requirement, rather than inferring geography from a provider’s headquarters.
OpenRouter’s Sensitive Info guardrail adds a useful layer before forwarding, with regex presets for emails, phone numbers, SSNs, credit cards, IPs, and provider-prefixed API keys, redact-or-block actions, and ReDoS-validated custom patterns. Read the fine print, though. It scans inputs only, never model responses, and its NLP-based name and address presets fail open on timeout. It shrinks the blast radius without closing the channel.
Seed your own canary tokens
You can adopt the researcher’s instrument yourself. Seeding a fake credential or a unique URL into the context your agents and pipelines already send costs almost nothing, and a free Thinkst Canarytoken is enough to start. Any touch of it becomes an incident with a timestamp and a source address. For AWS, canary keys in empty, fully logged accounts turn CloudTrail into a tripwire for prompt leakage, and the ARTEX user agents from this study are free detection content.
Beyond instrumentation, Moore’s advice comes as a ranked list. Host the model yourself if you can afford it, since that is the only arrangement where nobody else reads your prompts. Otherwise go directly to a first-tier provider, where you have one company, one contract, and one privacy policy to read. If you use routers, pick the upstreams yourself, verify that your privacy settings actually change where requests go, and stay away from reseller marketplaces whose capacity sometimes comes from stolen keys. The reader discussion after the disclosure, including Ars Technica readers asking whether leaked prompts could expose trade secrets or affect the patentability of unpublished inventions, suggests the audience for this advice reaches well beyond security teams.
Why a ZDR label is not a privacy boundary
OpenRouter’s email places the failure at MiniMax and the forwarders, and the evidence published so far fits that. OpenRouter’s ZDR machinery did what it was built to do. It filtered endpoints by declared policy, and the declarations did not describe what happened after the first hop. Novita’s trust center promises zero retention while its sub-processor list acknowledges model providers receiving inputs and outputs. AtlasCloud’s docs describe aggregation openly, while its OpenRouter endpoints carried a retention label that said nothing about the onward journey. Each party’s paperwork was, within its own frame, more or less honest. The customer’s mental model of one request, one provider, and one policy was the part that was wrong.
As inference becomes a commodity routed across a graph of resellers, gateways, and labs, the privacy of a call equals the privacy of its weakest hop, and no checkbox at the edge can see the whole graph. Closing that gap takes verifiable claims: sub-processor disclosure at endpoint granularity, contractual bans on undisclosed forwarding with audit rights that mean something, continuous independent canary auditing of the kind Dragos just ran at scale, and eventually technical attestation of where and how a request was processed. Some of that is achievable now with contracts and instrumentation. Some of it needs infrastructure the industry has not built.
Until then, the safe assumption for anyone sending sensitive data through an AI gateway is the one the canaries support: a prompt is a disclosure to every party in the chain, retained for periods nobody has published and readable by people you have never met.
Frequently asked questions
What did OpenRouter’s MiniMax email say?
OpenRouter told customers that requests sent to MiniMax models between June and October 2026 may have been read by a third party. A researcher’s unique test URLs inside those prompts were opened 40 to 130 hours after the requests were sent, including when providers forwarded requests to MiniMax. OpenRouter removed the ZDR label from the affected endpoints while it investigates.
Were credentials stolen on the MiniMax path?
No evidence of that has been published. None of the researcher’s AWS test keys were used on the MiniMax path. The canaries show that prompt content was stored and later accessed, and the email extends the warning to outputs.
Does OpenRouter’s ZDR flag guarantee zero data retention?
The zdr: true flag restricts routing to endpoints whose declared policy is Zero Data Retention. OpenRouter says that declaration is an attestation about the provider’s agreement with its upstream, and the platform does not independently verify the upstream’s contractual behavior. A request forwarded to another company can fall outside the label.
How can I check which OpenRouter endpoints qualify as ZDR?
OpenRouter publishes the eligible endpoints programmatically at /api/v1/endpoints/zdr, and the list updates as provider policies change.
How do I detect prompt leakage in my own AI pipelines?
Seed unique canary tokens, such as a free Thinkst Canarytoken, a fake AWS key in an empty and fully logged account, or a unique URL, into the context your agents send. Any later touch of the token is a timestamped sign that the prompt was retained or read.







