News
Attackers copied prompts inside three AI hosts for months.
Dragos says attackers copied customer prompts from inside three AI inference providers for months, including requests marked so they should never be stored, and none of the three knew until researchers told them.

Impact
- What was impacted
- Customer prompts, and secrets inside them, handled by third-party AI inference providers. Dragos says attackers were inside three providers, the earliest since mid-July, and copied traffic that included real customer requests as well as planted bait. Some stolen requests were marked Zero Data Retention, the setting that is supposed to mean a prompt is never stored. A separate path needed no break-in: sellers on an inference marketplace turned on request logging and collected what customers sent. OpenRouter, the gateway that routes requests to a pool of providers, said the research showed no breach of OpenRouter itself.
- Who was impacted
- Teams that send prompts, API keys, passwords, or private repository names through OpenRouter or straight to an inference host. Three providers have posted notices. Phala says to treat as exposed every request to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway from October 21 at 00:00 UTC through November 8 at 05:35 UTC. Together AI says stored samples from June and July requests to MiniMax M3, GLM-5.2, and Kimi K2.6 were downloaded. Novita says a routing mistake sent some MiniMax requests on to MiniMax's official API. Dragos has not named the three breached providers. OpenRouter has not named the three it removed. Neither has said whether those lists are the same three.
- What you need to do
- If a prompt in the Phala window contained a key, token, or password, rotate it. Phala keeps that whole window even though the canary credentials later misused were only the ones sent on October 21 and 22, because those tests sampled traffic rather than watching every request. If a June or July prompt to MiniMax M3, GLM-5.2, or Kimi K2.6 on Together AI contained a secret, rotate it: an unauthorized party downloaded the samples and tested credentials found in them. Together AI says it understands that copy was deleted and has no reason to believe the data went further. Do not treat a Zero Data Retention flag as proof a prompt stayed private after it left your network. Pick upstream providers yourself, check that a privacy setting actually changes where a request goes, and stay away from reseller marketplaces. A fake credential in the context your agents already send is a tripwire if any later use of it is treated as an incident. Dragos says two user-agent strings, artex-agent and artex-judge-r121, were not on a public indicator list it could find.
Attackers copied data from inside three AI inference providers for months. Dragos said none of the three knew until researchers told them. The earliest breach began in mid-July. The planted prompts were bait, processed alongside real customer requests, and the firm says those real requests were almost certainly copied too. Some of the stolen requests carried a Zero Data Retention mark. OpenRouter said it removed three providers from its platform, removed 125 endpoints from eligibility for that setting, and notified customers who might have been affected. It said its own systems were not breached. The 125 endpoints cover providers the researchers flagged and others removed as a precaution. An endpoint taken off that list needs assurances and a check before the status is restored. OpenRouter is adapting the researchers' checks so they can run on an ongoing basis, and it is tracing network paths to see where requests and responses may be forwarded.
The research seeded about two million prompts with unique web links and credentials between July and November, then watched for anyone using them. A planted secret can be learned only from the prompt that carried it. In the largest of the three breaches, a couple of hundred planted AWS keys each went out in exactly one request to one provider, during a few short windows in late October. Every request sent in those windows was stolen. Some of them were requests the provider had rejected before a model ever ran, which the firm reads as copying at the first point of arrival rather than deeper in the inference stack. The thief tested keys within hours in some cases, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock API keys and add IAM users, driving tens of thousands of AWS API calls. One attacker appears to have been inside two of the providers at once, using secrets from both from a single rented server. On November 1 that attacker brought in ARTEX, an AI penetration-testing framework, and used it only to confirm that keys it already held were still live.
A second finding needed no intrusion. Sellers on an inference marketplace listed models at the lowest prices, turned on request logging, and collected what flowed through. The people doing it looked like individuals on home connections. One, working from a Mac on home broadband in France, took passwords from two prompts, typed root SSH login attempts by hand, and used the same machine to test several planted AWS keys. One of those keys had passed through a different router that quietly forwarded some of its traffic to the marketplace, a hop no customer was told about. That router has since told the firm it stopped. Tracing about 2,000 paths found several hundred unidentified systems, and one seller that quietly adds about 1,000 hidden tokens to every prompt.
A Chinese AI lab, which the research does not name, fetched the secret links from more than half of the seeded prompts sent to its models between July and October, including about two-thirds of the requests marked for no data collection or Zero Data Retention. Prompts sent to U.S. hosting providers that resell that lab's models triggered the same collection and flowed back to the lab. A git client on another cloud later tried to clone private repositories named in the prompts, and only for that lab's models. One large U.S. provider was still requesting a secret link more than two months after a single request to its search endpoint. Twenty-four days after one prompt was sent, its secret link was posted into a chat on Feishu, the workplace app known as Lark elsewhere. It was opened within two minutes from three desktops and three phones in Beijing and Shanghai. The firm had published only the first eight characters of that provider's link, so whoever posted it was reading a stored copy of the full prompt.
Three providers have posted public notices. Phala said an attacker took control of an inference server and could view request data while it was processed. Phala contained that on November 8 and restored the model service with added protections. Together AI said an AI coding agent committed a credential to the company's public GitHub repository, after which an unauthorized party downloaded stored prompt samples kept for routine software testing. Five of the credentials that party tested were canaries from the experiment. Together AI has made zero data retention the default for new accounts. Novita said a routing configuration sent some MiniMax requests on to MiniMax's official API, beyond what those customers intended. The provider behind the mid-July breach, which Dragos did not name, first blamed an internal misconfiguration, then a partner's data processing, and later found that attackers had compromised its internal credentials. It had not disclosed the incident publicly as of this November 22 report. In October, Railway said an intrusion it suffered was likely part of a series of attacks aimed at platform companies in order to reach AI inference providers. The researchers are working with about a dozen providers, and as of this report planted keys and secret links were still being used months after the first thefts.
For your AI agent
Agent instructions — copy paste to Claude
You are helping me respond to prompt theft at third-party AI inference providers, reported by Dragos. Attackers copied customer prompts from inside three providers for months, including some requests marked Zero Data Retention. OpenRouter says its own systems were not breached. Do not write or send a prompt, a canary, or a request in order to test whether a provider is still copying traffic. Do not try to use a stolen key, list someone else's buckets, or reproduce the AWS activity described in the report. Ask before rotating a credential, disconnecting a provider, changing a privacy setting, or deleting a log. Facts from the report: - Dragos said attackers were inside three AI inference providers and none of the three knew until researchers told them. The earliest breach began in mid-July. Planted prompts were bait, processed next to real customer requests, and the firm says those real requests were almost certainly copied too. - Some stolen requests were marked Zero Data Retention, the setting that is supposed to mean a prompt is never stored. A privacy flag binds only the first hop. Do not tell me a flag proves a prompt was not copied. - OpenRouter routes customer requests to a pool of third-party providers. It said it removed three providers, removed 125 endpoints from Zero Data Retention eligibility, and notified potentially affected customers. The 125 endpoints cover providers the researchers flagged plus others cut as a precaution. Any endpoint removed in the review needs assurances and validation before that status returns. OpenRouter said the findings showed no breach of OpenRouter itself. It is adapting the researchers' checks so they can run continuously and is verifying network paths for forwarding to third parties. - Dragos has not named the three breached providers. OpenRouter has not named the three it removed. Neither has said publicly whether they are the same three. Do not decide that a named company is one of those three unless that company said so. - The research seeded about two million prompts with unique web links and credentials from July through November. A planted secret can be learned only from the prompt that carried it. - In the largest breach, a couple of hundred planted AWS keys were each sent in exactly one request to one provider during a few short windows in late October. Every request in those windows was stolen, including some the provider rejected before a model ran. The firm reads that as copying at the first arrival point. The thief tested keys within hours in some cases, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock API keys and add IAM users, with tens of thousands of AWS API calls. Do not repeat those calls. - One attacker appears to have compromised two of the providers at once and used secrets from both from a single rented server. On November 1 the same attacker used ARTEX, an AI penetration-testing framework, only to check that keys it already held were still live. The user agents artex-agent and artex-judge-r121 were not on a public indicator list Dragos could find. Treat them as detection strings from this report, not as proof by themselves. - A second finding needed no break-in. Marketplace sellers listed models at the lowest prices, enabled request logging, and collected the traffic. One person, on a Mac on home broadband in France, took passwords from two prompts, typed root SSH login attempts by hand, and tested several planted AWS keys from that machine. One key passed through a different router that quietly forwarded some traffic to the marketplace. Customers were not told. That router later told the firm it stopped. Tracing about 2,000 paths found several hundred unidentified systems, and one seller that adds about 1,000 hidden tokens to every prompt. Do not invent the seller's name or the router's name. - An unnamed Chinese lab fetched the secret links from more than half of the seeded prompts sent to its models between July and October, including about two-thirds marked no data collection or Zero Data Retention. Prompts sent to U.S. hosts that resell that lab's models triggered the same collection. A git client on another cloud later tried to clone private repositories named in prompts, only for that lab's models. One large U.S. provider kept requesting a secret link more than two months after a single request to its search endpoint. Do not guess which lab or which U.S. provider. - Twenty-four days after one prompt, its secret link was posted in a Feishu chat (Lark outside China) and opened within two minutes from three desktops and three phones in Beijing and Shanghai. Only the first eight characters of that provider link had been published, so the poster was reading a stored full prompt. - Phala posted a preliminary incident report: an attacker took control of an inference server and could view request data while it was processed. Contained on November 8. The model service was restored with added protections. Treat as exposed every request to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway from October 21 at 00:00 UTC through November 8 at 05:35 UTC, and rotate every key, token, password, or other secret in those prompts. Canary credentials had been planted since July 1, and only the ones sent on October 21 and 22 were later misused. The tests sampled traffic, so Phala keeps the wider window. Do not shrink that window. - Together AI said an unauthorized party downloaded stored samples of customer prompts after an AI coding agent committed a credential to Together AI's public GitHub repository. The samples were for routine testing and came from June and July requests to MiniMax M3, GLM-5.2, and Kimi K2.6. The party searched the samples for credentials to other services and tested them. Five tested credentials were canaries. Together AI said it has no reason to believe the data went further, understands the copy was deleted, and made zero data retention the default for new accounts. A secret that was inside those samples should still be rotated. Do not treat "deleted" as "never tested." - Novita said a routing configuration sent some MiniMax requests on to MiniMax's official API, beyond what those customers intended. Describe that as a routing disclosure, not as one of the three intrusions, unless I show you a notice that says otherwise. - The unnamed provider behind the mid-July breach first blamed an internal misconfiguration, then a partner's data processing, then found attackers had compromised its internal credentials. It had not posted a public notice as of this November 22 report. - Railway said in October that an intrusion it suffered was likely part of attacks aimed at platform companies in order to reach inference providers. Do not turn that into a claim that Railway is one of the three providers. - Researchers are working with about a dozen providers. As of this report, planted keys and secret links were still being accessed months after the first thefts. - Advice in the research: self-hosting is the only option in which no third party reads a prompt. Going direct to one provider leaves one company, one contract, and one privacy policy to check. On a router, pick the upstream providers yourself, verify that privacy settings change where requests go, and stay away from reseller marketplaces, some of whose capacity the research says runs on stolen keys. A fake credential in context the agents already send works as a tripwire. The researchers said a free Thinkst Canarytoken is enough to start. Scrubbing known credential formats from prompts helps and still misses sensitive data that is not shaped like a key. Do not present any of that as a patch for a provider I do not control. Context: - What was impacted: prompts and secrets at third-party inference providers, including some marked Zero Data Retention. OpenRouter itself was not reported breached. - Who should care: anyone who sent prompts or secrets through OpenRouter or directly to these hosts, especially the Phala DeepSeek window and the Together AI June–July samples named above. - What to do: rotate secrets that were inside the exposed windows, stop assuming a privacy flag protected a prompt after it left our network, and choose upstream providers explicitly. Do this in order. Ask before rotating a secret, removing a provider, or changing a setting. Prefer read-only inventory first. Do not write or run an exploit, a proof of concept, or a live test against a provider. 1. Inventory - List where we send prompts: OpenRouter, a named provider, a marketplace, or a self-hosted model. If you cannot see that configuration, say so and stop. - Record whether any route uses a default provider pool, a reseller marketplace, or a privacy flag such as Zero Data Retention. A flag is not evidence the prompt stayed private. - Record whether we sent requests to DeepSeek V4 Flash 0731 through Phala between October 21 at 00:00 UTC and November 8 at 05:35 UTC, and whether June or July requests to MiniMax M3, GLM-5.2, or Kimi K2.6 went to Together AI. If the logs do not say, say the window is unknown. Do not mark an unknown route as safe. - Note Novita only if we sent MiniMax requests through it. That notice is a routing issue to MiniMax's official API. - If logs show the user agents artex-agent or artex-judge-r121, record the time and the account. Do not treat a missing log line as proof nothing was taken. 2. Contain - Do not tell me to disconnect every provider. Propose pausing routes that still use a default pool or a reseller marketplace, and ask before changing them. - If a prompt in a named window contained a live secret, say which secret and which window. Ask before rotating it. 3. Remediate - For the Phala window, plan to rotate every key, token, password, or other secret that was inside those prompts. Ask before changing one. - For Together AI samples from those three models in June and July, plan to rotate secrets that were in the prompts. State Together AI's claim that the downloaded copy was deleted and that it has no reason to believe the data went further, and still recommend rotation for secrets that were in the samples. - Where we use a router, plan to pin the upstream providers ourselves and to check that a privacy setting changes the actual route. Ask before editing that configuration. - Do not invent a provider patch version. This report does not give one. - A canary credential is optional and only if I ask for one. Do not send it to a provider as a test from this task. 4. Verify - After a rotation or route change I asked you to make, read that setting or that credential's status again. - If I did not ask you to change anything, report the inventory only. 5. Report back - Which prompt routes you could see, which named windows might include our traffic, which secrets you recommend rotating, what you changed, and that Dragos has not named the three breached providers. Do not invent CVEs, provider names, victim counts, or a claim that a particular prompt was stolen. If a check needs access I have not given you, say what you need and stop.