OpenRouter Cuts 3 AI Providers Over Prompt Theft: What to Do
TL;DR · the short version
On 22 November 2026 OpenRouter said it had off-boarded three inference providers, removed 125 endpoints from Zero Data Retention (ZDR) eligibility and notified potentially affected customers, after security researchers found that some providers "did not uphold their data policies". OpenRouter says there was no breach of OpenRouter itself. The research is Dragos's (HD Moore, 19 November): nearly two million prompts seeded with unique links and credentials from July to November 2026 found at least three AI providers breached for months without knowing, some ZDR-marked requests stolen, and planted AWS keys tested within hours. Phala, Together AI and Novita have published their own notices. Nobody has named the three off-boarded or the three breached providers, or said they are the same. Rotate any secret that was in a routed prompt, pin upstream providers, verify ZDR routing and plant your own canaries.
OpenRouter has off-boarded three AI inference providers and removed 125 endpoints from Zero Data Retention (ZDR) eligibility, it said on 22 November 2026, after security researchers found that some providers had data compromised or "did not uphold their data policies". OpenRouter says it has notified all potentially affected customers and that the findings "did not show any breach of OpenRouter". The research behind it, published by Dragos on 19 November, says at least three inference providers were breached for months without knowing, that some stolen requests were marked ZDR, and that planted AWS keys were being tested by an attacker within hours. Neither OpenRouter nor Dragos names the providers.
Who this is for: security leaders, platform and AI engineering teams, and anyone whose apps, agents or staff send prompts through OpenRouter, another LLM router or an inference marketplace.
Independent analysis based on OpenRouter's announcement, Dragos's research post, and the notices published by Phala, Together AI, Novita and Railway, all read on 22 November 2026. The story was first brought to us by The Hacker News's report of the same day. PlayCISO is not affiliated with any company named. Facts may change as providers publish more.
Confirmed vs not yet known
OpenRouter says: it off-boarded three providers; removed 125 endpoints from ZDR eligibility (providers the researchers flagged plus others "as a precaution"); notified all potentially affected customers; found no breach of OpenRouter; and is adapting the researchers' tests to run continuously, verifying network paths for forwarded requests, and requiring assurances and validation before restoring ZDR eligibility.
Providers say (their own notices): Phala: an attacker "took control of the inference VM" serving DeepSeek V4 Flash 0731. Together AI: an unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential. Novita: a routing configuration issue sent some MiniMax requests to MiniMax's official API.
Dragos says (research, not a provider admission): at least three providers breached, the earliest from mid-July; none knew until Dragos told them; the largest involved "a couple hundred" planted AWS keys in late October, and every request sent in those windows was stolen.
Not yet known: which three providers OpenRouter removed; which three providers Dragos found breached; whether those are the same three; whether Phala and Together AI are among either group (no source says so); who the attackers are; how many customers or requests were affected. OpenRouter says not every affected provider has posted publicly.
What OpenRouter announced
OpenRouter is an LLM router: one API that sends each request to one of many inference providers. On 22 November 2026 it published "Data Processing Incidents on Inference Providers". It says security researchers reported that certain providers "had data compromised or did not uphold their data policies", that it has been working with the researchers and the providers, and that it has:
- off-boarded three providers;
- removed 125 endpoints from ZDR eligibility, including providers the researchers flagged and others removed as a precaution;
- notified all potentially affected customers, including where the provider has not posted publicly.
It also lists ongoing work: running the researchers' style of testing continuously, verifying the network paths of requests or responses forwarded to third parties, and requiring assurances and validation before any removed endpoint regains ZDR eligibility. The announcement links three provider notices, from Novita, Phala and Together AI, and the Dragos research. It does not say which three providers were off-boarded.
What the Dragos canary research found
Dragos's post, "Three breaches, endless leaks: Your AI supply chain needs canaries", is by HD Moore, Dragos's chief research officer. Dragos says it sent "nearly two million AI prompts containing unique, trackable secrets" (unique links and fake credentials, known as canaries) to AI providers from July to November 2026, so that any later use of a secret shows who read the prompt.
Three providers breached for months
- Data from "at least three AI providers" was stolen. "The earliest incident dates back to the middle of July," and "none of the three providers knew about their breach until we notified them."
- "The largest breach involved a couple hundred of our AWS keys." Each key went in exactly one request to one provider, in a few short windows in late October, and Dragos says "every request we sent during those windows was stolen", including requests the provider rejected before a model ever ran. Dragos concludes the attacker was copying requests "at the provider's front door".
- In the other two breaches, Dragos says, "some of the stolen requests were marked as ZDR."
- The attacker tested keys for liveness, in some cases within a few hours, then listed S3 buckets, searched Secrets Manager and Parameter Store, tried to enable Bedrock models and create Bedrock API keys, and added their own IAM users to keep access. Dragos counts "tens of thousands of AWS API calls" against its bait keys; one key seeded on 21 October was first used the next day and still in use three weeks later.
- "One adversary appears to have compromised two of these providers at the same time": secrets sent to both were used together, almost immediately, from a single rented server. Its tooling moved from a script posing as Chrome on Windows, to Boto3, to the AWS CLI on Kali Linux.
- On 1 November the same attacker added ARTEX, an LLM-driven penetration testing framework, and used it only to make "a handful of calls" confirming keys were still live. Dragos names the CloudTrail user agents
artex-agentandartex-judge-r121, says it found neither in any public indicator list, and warns that self-identification appeared "only in a small fraction of traffic".
Marketplaces, routers and a Chinese lab
- Marketplace sellers: on an inference marketplace, Dragos says sellers "only had to list a model at the lowest price, enable request logging, and wait". People who bought from them opened secret links and tried stolen passwords against logins and SSH.
- A router forwarding to the marketplace: Dragos says one router "was quietly forwarding some of its traffic to the marketplace", and stopped after being told.
- Hidden hops: of about two thousand routes Dragos traced, "several hundred" passed through a system it could not identify, and "one marketplace seller quietly adds about a thousand hidden tokens to every prompt".
- An unnamed Chinese AI lab: its systems fetched the secret links from "more than half of the prompts" sent to its models from July to October, including "roughly two-thirds of the requests marked as 'no data collection' or Zero Data Retention", usually two or three days later. Dragos also says that 24 days after one prompt was sent, its secret link was posted into a Feishu chat channel and opened within two minutes from devices in Beijing and Shanghai. We covered the matching OpenRouter notice about MiniMax models in OpenRouter MiniMax prompts may have been read; neither Dragos nor OpenRouter names the lab.
What each provider says
Our Breach Index tracks each of them with a dated, sourced timeline: Phala, Together AI and OpenRouter (not a breach, by OpenRouter's own account).
Phala: attacker controlled an inference server
Phala's preliminary incident report (10 November 2026) says an attacker chained a legacy internet-facing service with management interfaces that "did not independently authenticate callers", created unauthorized confidential VMs and obtained application credentials. Separately, a DeepSeek V4 Flash 0731 deployment on a new GPU server, serving production traffic "without our full production security hardening", was compromised: the attacker "took control of the inference VM, making in-flight request data visible during processing." Phala says it served this model only through the route OpenRouter sent it, and asks users to treat DeepSeek V4 Flash 0731 requests from 21 October 2026 00:00 UTC to 8 November 2026 05:35 UTC as exposed (canary-confirmed exposure: 21 to 22 October) and to "treat any credentials or other sensitive information in those requests as compromised." It lists containment on 5 November, a KMS root key rotation on 6 November, and completion of model-incident containment at 05:35 UTC on 8 November.
Useful next, if this was relevant to you
Together AI: stored prompt samples downloaded
Together AI's security notice (21 November 2026, updated 22 November) says prompt samples kept for routine software testing were stored in a private Hugging Face repository, and that "an AI coding agent committed a credential that provided access to the private Hugging Face repository" while the team was developing a public GitHub repository. Automated scanning caught it and the commit was reverted, but an unauthorized party used the credential to download the samples, searched them for other credentials and tested them; Together AI says five of those were canaries from the Dragos experiment. The samples came from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, "primarily on June 24, July 4, July 8, and July 14, 2026." Together AI says it is working with an external forensic firm and law enforcement, has found no other prompt leakage, has made ZDR the default for new accounts, and has tightened credential scoping, pull-request secret checks and its processes for AI coding agents. It advises customers with sensitive prompts to enable ZDR.
Novita: MiniMax requests forwarded to MiniMax
Novita's privacy notice (22 November 2026) says "a routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly", so some MiniMax requests from 28 May to 11 November 2026, direct and via OpenRouter, went to MiniMax's official API instead of Novita's own infrastructure, and "your request contents may have been accessed or retained by a third party." It says the researcher saw accesses to embedded URLs and hostnames but no use of the AWS test keys, that the pattern suggests automated content review (still being checked with MiniMax), that it suspended the affected models, and that users should rotate any secrets those requests contained.
Railway: what it actually said
The Hacker News reports that Railway said in October its intrusion was likely part of attacks on platform companies aimed at reaching inference providers. Railway's own words are a little different. Its preliminary report (22 October 2026) says the attack "was attempting to obtain AI provider API keys to perform distillation attacks". Its full incident report (10 November) says "we assess this to be one of a series of attacks on platforms that hold AI keys" and that from 21 October it sent affected providers lists of potentially exposed keys. No source links Railway's intrusion to the three provider breaches Dragos describes. See Black October: downstream breaches for the wider October wave.
Timeline
| Date (2026) | Event | Source |
|---|---|---|
| 28 May to 11 Nov | Some MiniMax requests sent through Novita forwarded to MiniMax's official API by a routing configuration issue | Novita privacy notice |
| 24 Jun, 4, 8, 14 Jul | Main dates of the requests whose stored prompt samples were later downloaded | Together AI security notice |
| Mid-July | Earliest of the three provider breaches Dragos found | Dragos |
| July to November | Dragos sends nearly two million canary prompts to AI providers | Dragos |
| 20 to 22 Oct | Railway attack window; Railway says the attacker sought AI provider API keys | Railway preliminary report |
| 21 Oct, 00:00 UTC | Start of Phala's precautionary exposure window for DeepSeek V4 Flash 0731 (canary-confirmed exposure 21 to 22 Oct) | Phala preliminary report |
| Late October | Largest breach: a couple hundred planted AWS keys stolen; keys tested within hours | Dragos |
| 1 Nov | Same attacker uses ARTEX to check stolen keys (artex-agent, artex-judge-r121) | Dragos |
| 5 to 8 Nov | Phala contains the intrusion (5 Nov), rotates its KMS root key (6 Nov), completes model-incident containment (8 Nov, 05:35 UTC) | Phala preliminary report |
| 10 Nov | Phala publishes its preliminary incident report; Railway publishes its full incident report | Phala; Railway |
| 19 Nov | Dragos publishes the canary research | Dragos |
| 21 Nov | Together AI publishes its security notice (updated 22 Nov) | Together AI |
| 22 Nov | Novita publishes its privacy notice; OpenRouter announces it off-boarded three providers and removed 125 endpoints from ZDR eligibility | Novita; OpenRouter |
Why this matters: a privacy label is not a control
When you send a prompt through a router, the router, the provider it picks, and anyone that provider forwards to can all see it in clear text. Those parties are subprocessors of your data, and some of them you never chose. Dragos's findings show three ways that goes wrong: the provider itself is breached (ZDR does not stop an intruder copying requests at the front door), a seller deliberately logs traffic, or a hop you did not know about forwards the request on. OpenRouter pulling ZDR from 125 endpoints is a public admission that the label could not be relied on for those endpoints.
The other lesson is speed. Dragos says the attacker tested stolen AWS keys within hours. Any secret that sits in a prompt, a system prompt or an agent's context should be treated as something an intruder at any hop can use the same day.
What security leaders should do this week
- Inventory where your AI traffic goes. List every app, agent, internal tool, CI job and developer machine that calls OpenRouter, another router or an inference marketplace, including personal keys and coding agents configured with a router base URL. Search code and config for router hosts and keys.
- Read your notices and check your logs. OpenRouter says it notified all potentially affected customers. Pull your router activity logs for the windows the providers give: Phala's DeepSeek V4 Flash 0731 (21 October 00:00 UTC to 8 November 05:35 UTC), Novita's MiniMax requests (28 May to 11 November) and Together AI's MiniMax M3, GLM-5.2 and Kimi K2.6 requests (June and July).
- Rotate any secret that was ever in a routed prompt during those windows. API keys, cloud keys, database URLs, tokens, passwords and signed URLs, in dependency order, then review the accounts for misuse. Phala and Novita both tell users to rotate. Use the leaked credential response steps and the LLM API key exposure playbook.
- Pin upstream providers. Dragos recommends choosing upstream providers yourself rather than letting a router pick, and going direct to a first-tier provider where you can. On OpenRouter that means per-request
provider.only(an allow-list) withallow_fallbacks: falsefor sensitive workloads, rather than a block-list. - Verify ZDR routing, do not assume it. Set
zdr: trueanddata_collection: "deny"per request or account-wide, then check that they actually change which endpoints serve your requests. Dragos's advice is to "verify that privacy settings change where requests go". Avoid reseller marketplaces for anything sensitive. - Plant canary tokens in agent and pipeline context. Dragos recommends adding a fake credential to agent and pipeline context and treating any use of it as an incident. It is cheap, and it is how all of this was found.
- Strip credentials from prompts. Give agents short-lived, scoped credentials through tools rather than pasting secrets into context, and scan prompts and context for secrets before they leave your network (the secret scanner shows the patterns). Dragos calls secret filtering helpful but "not a panacea".
- Add the ARTEX user agents to cloud detections. Dragos suggests watching CloudTrail for
artex-agentandartex-judge-r121. Alert on them, but do not rely on them: Dragos says self-identifying user agents were only a small fraction of the attacker's traffic, so also alert on new IAM users, Bedrock model enablement and Bedrock API key creation, and Secrets Manager and Parameter Store reads from unfamiliar origins. - Treat routers and providers as subprocessors. Ask each one who else processes your requests, where, for how long, and how you will be told when that changes. Score them with the vendor risk scorecard.
Put it to work
- AI Subprocessor Impact Lab: practise tracing an AI vendor change to the subprocessors that actually handle your prompts.
- Leaked Credential Response: what to do, in order, when a key may be in someone else's hands.
- Secret Rotation Order Planner: plan the order to rotate secrets that may have been in prompts.
- AI Footprint: map where AI tools and services are in use across the organisation.
FAQ
Was OpenRouter hacked?
OpenRouter says no. Its 22 November 2026 announcement says the researchers' findings "did not show any breach of OpenRouter" or unauthorized access to its systems. The problems were at inference providers that OpenRouter routes requests to. That is OpenRouter's own assessment; no independent investigation of OpenRouter has been published.
Which AI providers were breached?
Dragos says data from "at least three AI providers" was stolen, but does not name them, and OpenRouter does not name the three providers it off-boarded. Two providers have published their own incident notices: Phala says an attacker took control of an inference server running DeepSeek V4 Flash 0731 (precautionary window 21 October 00:00 UTC to 8 November 05:35 UTC), and Together AI says an unauthorized party downloaded stored prompt samples from June and July 2026 after an AI coding agent committed a credential. Novita published a separate privacy notice about MiniMax requests forwarded to MiniMax's own API. No source says which, if any, of these are the three Dragos breaches or the three providers OpenRouter removed.
Does Zero Data Retention protect my prompts?
Not on its own. Dragos says some of the stolen requests were marked as ZDR and that "a privacy flag doesn't help when the attacker is already inside"; it also says a privacy flag "only binds the first hop". OpenRouter removed 125 endpoints from ZDR eligibility and says it will require assurances and validation before restoring any. Keep using ZDR and data-collection settings, but treat them as a policy to verify, and keep secrets out of prompts.
What should I rotate?
Any credential, token, password or signed URL that was in a prompt or in an agent's context sent through an affected route. Phala says to treat credentials in DeepSeek V4 Flash 0731 requests routed to it by OpenRouter between 21 October 00:00 UTC and 8 November 05:35 UTC as compromised. Novita says to rotate secrets in MiniMax requests sent through it from 28 May to 11 November 2026. If OpenRouter or a provider notified your account, use the dates in that notice. If you cannot tell which provider served a request, rotate anyway: Dragos says attackers tested stolen AWS keys within hours.
Who did the canary research?
Dragos. The post "Three breaches, endless leaks: Your AI supply chain needs canaries" (19 November 2026) is by HD Moore, Dragos's chief research officer. It says the team sent nearly two million prompts containing unique, trackable secrets to AI providers between July and November 2026.
What are the artex-agent and artex-judge-r121 user agents?
Dragos says that on 1 November 2026 the attacker behind the largest breach added ARTEX, an LLM-driven penetration testing framework, and used it only to make "a handful of calls" confirming that stolen AWS keys were still live. Those calls carried the user agents "artex-agent" and "artex-judge-r121" in AWS CloudTrail, which Dragos says it did not find in any public indicator list. Dragos also notes that self-identifying user agents appeared in only a small fraction of the attacker's traffic, so a match is a useful signal but a miss proves nothing.
Sources
- OpenRouter: Data Processing Incidents on Inference Providers (22 Nov 2026)
- Dragos (HD Moore): Three breaches, endless leaks: Your AI supply chain needs canaries (19 Nov 2026)
- Phala: November preliminary incident report (10 Nov 2026)
- Together AI: Security event, prompt leakage for requests in June and July of 2026 (21 Nov 2026, updated 22 Nov)
- Novita: Privacy notice regarding MiniMax requests on Novita (22 Nov 2026)
- Railway: Preliminary report, Oct 21st security incident (22 Oct 2026) and Railway: incident report (10 Nov 2026)
- OpenRouter documentation: provider selection and Zero Data Retention
- The Hacker News: OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms (22 Nov 2026): the report that brought this to us; facts above are taken from the primary sources
This article is educational guidance for security and engineering leaders, not legal advice. Check the notice sent to your own account and any further statements from OpenRouter and the providers before acting on the details.