Your prompts were supposed to be private. For months, inside at least three AI inference providers, they were being copied — including the ones marked Zero Data Retention, and including requests that were rejected before a model ever ran.
That is the finding from industrial cybersecurity company Dragos, published November 19, 2026. OpenRouter, the gateway that routes customer AI requests to a pool of third-party providers, responded today: it has removed three providers from its platform, removed 125 endpoints from eligibility for Zero Data Retention, and notified all potentially affected customers. The findings showed no breach of OpenRouter itself, the company said.
The honest framing up front: I haven’t tested any of this hands-on, and there is one big thing nobody has confirmed. The three providers Dragos caught and the three providers OpenRouter removed have not been confirmed to be the same three. Dragos has not named the providers it found breached, and OpenRouter has not named the providers it removed. Three providers have posted public notices of their own — Phala, Together AI and Novita — and those notices are worth reading in full if you used any of them. Everything below is the research and the notices as reported, plus what you should do now.
How the theft was caught
Dragos seeded about two million prompts with unique web links and credentials — canary tokens — between July and November 2026, then watched for anyone using them. A planted secret can only be learned in one place, so any use of it means someone read the prompt that carried it.
The earliest of the three breaches began in mid-July. The planted prompts were bait, but they were processed alongside real customer requests, and those were almost certainly copied too, the firm said.
The largest of the three breaches involved a couple hundred planted AWS keys, each sent in exactly one request to one provider during a few short windows in late October. Every request sent during those windows was stolen, and some came from requests the provider had rejected before a model ever ran — indicating the attacker was copying traffic 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 to keep access, driving tens of thousands of AWS API calls. One attacker appears to have compromised two of the providers at once, using secrets from both from a single rented server. On November 1, the same attacker brought in ARTEX, an AI-driven penetration testing framework, and used it only to confirm that keys it already held were still live.
Planted keys and secret links were still being accessed months after the first thefts, as of this week. The researchers are now working with about a dozen providers.
As Dragos chief research officer HD Moore put it: “A privacy flag doesn’t help when the attacker is already inside.”
The second finding: some marketplaces were collecting prompts on purpose
The second finding needed no break-in. Sellers on an AI inference marketplace listed models at the lowest prices, enabled request logging, and collected what flowed through, the research found.
The people doing it looked like individuals on home internet connections. One, working from a Mac on home broadband in France, took passwords from two prompts and typed root SSH login attempts by hand, then used the same machine to test several planted AWS keys. One of those keys had travelled through a different router that quietly forwarded some of its traffic to the marketplace, a hop that no customer was told about. Route tracing across about 2,000 paths found several hundred unidentified systems, plus 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 no data collection or Zero Data Retention. Prompts sent to U.S. hosting providers that resell the lab’s models triggered the same scraper, flowing back to the lab. A git client on another cloud later tried to clone private repositories referenced in the prompts, but only for that lab’s models. Twenty-four days after one prompt was sent, its secret link was posted into a chat on Feishu, the Chinese workplace app known as Lark elsewhere. The link was opened within two minutes from three desktops and three phones in Beijing and Shanghai. In practice, Moore wrote, the flagged prompts were kept, passed to other companies, and read by human beings.
What the three providers have said
Phala. An attacker took control of an inference server at Phala and could view request data while it was being processed, the provider said in a preliminary incident report last week. The incident was contained on November 8, and the model service has been restored with added protections. Requests to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway between October 21 at 00:00 UTC and November 8 at 05:35 UTC should be treated as exposed, the report says, and every key, token, password, or other secret within them should be rotated.
Together AI. An unauthorized party downloaded stored samples of customer prompts after an AI coding agent committed a credential to the company’s public GitHub repository, the company said in a security notice on November 21. The samples had been collected for routine software testing and were generated by requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models in June and July. The party searched the samples for credentials to other services and tested them, and five of the tested credentials were canaries from the researchers’ experiment. The company said it has no reason to believe the data went further, understands it has been deleted, and has made zero data retention the default for new accounts.
Novita. Novita confirmed this week in a privacy notice that a routing configuration issue caused some MiniMax requests to be sent on to MiniMax’s official API, beyond what was intended for those customers. The issue ran from May 28 to November 11.
One more honest note: the provider behind the mid-July breach first blamed an internal misconfiguration, then a partner’s data processing, before its own investigation found that attackers had compromised its internal credentials. It had still not disclosed the incident publicly as of this week. And Railway, the platform company, said in October that an intrusion it suffered was likely part of a series of attacks built to target platform companies with the end goal of compromising AI inference providers.
What OpenRouter is doing now
OpenRouter said in its incident announcement that the 125 endpoints removed from Zero Data Retention eligibility cover providers flagged by the researchers plus others cut as a precaution. Any endpoint removed during the review will require assurances and validation before its status is returned.
The gateway is adapting the researchers’ testing approach so that the checks can run on an ongoing basis, and is verifying network paths to identify where requests and responses may be forwarded to third parties. Not all affected providers have posted publicly, the gateway said, and it has notified affected users either way.
What to do now
If you used DeepSeek V4 Flash 0731 through OpenRouter between October 21 and November 8: treat those requests as exposed. Rotate every key, token, password or other secret that appeared in them. That is Phala’s own instruction, and it is the most concrete exposure window published so far.
If you used OpenRouter at all in the last few months: check your cloud logs — AWS CloudTrail and equivalents — for unexpected calls, new users, or new keys. The attacker in the largest breach listed buckets, searched secret stores and tried to create persistent access. A stolen key that still works is the thing to find first.
If you used Together AI, Novita or MiniMax models: read the providers’ notices and check whether your requests fall in the affected windows — June and July for the Together AI samples, May 28 to November 11 for Novita’s routing issue.
For every AI API you use: stop sending real secrets in prompts. The researchers’ advice starts from one limit: a privacy flag binds only the first hop, and AI agents will send whatever they can read, credentials included. Where a prompt goes is a choice, and a customer who does not choose hands the choice to the lowest price.
The practical defences, from the researchers themselves:
- Pick your upstream providers yourself instead of accepting a router’s default pool, and verify that privacy settings actually change where requests go.
- Stay away from reseller marketplaces — some of whose capacity, the research found, runs on stolen keys.
- Go direct to a first-party provider where it matters: one company, one contract, one privacy policy to check. Self-hosting is the only option in which no third party reads a prompt at all.
- Plant a fake credential as a tripwire in the context your agents and pipelines already send, and treat any use of it as an incident. A free Thinkst Canarytoken is enough to get started.
- Scrub known credential formats from prompts before they leave your systems — it helps, though it misses sensitive data that is not shaped like a key.
What we still don’t know
- Which three providers OpenRouter removed. Neither OpenRouter nor Dragos has named them, and neither has said publicly whether the removed three are the same as the breached three.
- How many real customer prompts were copied. The research proves the mechanism and the scale of the canary theft; the full customer impact is still being assessed.
- Whether the stolen data went further. Together AI says it has no reason to believe its samples went further; for the other breaches, that question is open.
- When removed endpoints get Zero Data Retention back. OpenRouter says assurances and validation are required first — no timeline published.
Our take
The headline is the provider removals, but the real story is the privacy flag. Zero Data Retention did not protect these prompts, because a flag binds only the first hop — it cannot help when the attacker is already inside the provider. Until routing and verification improve, assume any prompt can be read by whoever sits on its path, and keep real secrets out of it.