OpenRouter drops 3 providers after research reveals months of prompt theft; DeepSeek swapped for 53 hours

Research using nearly 2 million prompts seeded with trackable secrets revealed prompt theft at AI inference providers for months — and the providers themselves did not know they had been breached until the researchers told them. OpenRouter, the gateway that routes customer requests to multiple providers, announced on November 22 that it removed three providers from its platform and stripped 125 endpoints of eligibility for Zero Data Retention, the setting that promises a prompt is never stored. The company says OpenRouter itself was not breached, and that it notified all potentially affected customers.

The research comes from Dragos, an industrial cybersecurity company. Chief research officer HD Moore planted unique credentials and links — so-called canary tokens — in about 2 million prompts sent between July and November through routers such as OpenRouter and Vercel, resellers and hosts. When a planted secret was used, the researchers knew that prompt had been read — proof that the prompt theft was real. The disclosure was made on November 19, according to Ars Technica.

Three providers breached, and a marketplace that logs everything

The researchers found evidence of breaches at at least three hosting providers. In the largest case, a couple of hundred planted AWS keys — each sent in exactly one request, to a single provider, during short windows in late October — were reused tens of thousands of times in AWS API calls: the attackers listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock keys and IAM users. Every request sent during those windows was stolen, including some the provider rejected before any model ran — a sign the theft happened at the front door, not deep in the inference stack. One attacker appears to have compromised two providers at once, from a single rented server, and on November 1 used ARTEX, an AI-driven penetration-testing framework, to confirm the keys still worked. The researchers say real customer requests processed alongside the tests were almost certainly copied too — the prompt theft was not limited to the bait.

Another part of the research needed no break-in: sellers on an AI inference marketplace listed models at the lowest prices, enabled request logging, and collected whatever flowed through. One of them, on a home connection in France, took passwords from two prompts and tried root SSH logins by typing them in by hand. A Chinese AI lab, which the research does not name, fetched the secret links from more than half of the prompts sent to its models between July and October — including about two-thirds of those marked no data collection or Zero Data Retention. And a secret link sent to a US provider showed up 24 days later in a chat on Feishu, the Chinese workplace app.

Dragos has not named the three breached providers, and OpenRouter has not named the three it removed — nor has it been confirmed publicly whether they are the same three.

What the providers have disclosed

Three providers have disclosed incidents of their own, listed by OpenRouter. Phala said an attacker took control of an inference server and could view request data while it was being processed; the incident was contained on November 8. Requests to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway between October 21 (00:00 UTC) and November 8 (05:35 UTC) should be treated as exposed, and every key, token or password in them should be rotated.

Together AI said an unauthorized party downloaded stored samples of customer prompts after an AI coding agent published a credential in the company's public GitHub repository. The samples came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models in June and July; five of the credentials tested by the intruder were canaries from the research. The company says it has no reason to believe the data went further, and has made Zero Data Retention the default for new accounts.

Novita confirmed a routing configuration issue that sent some MiniMax model requests to MiniMax's official API — a case Geekiki already covered: OpenRouter warns that prompts sent to MiniMax models may have been read by third parties.

In the same week, a DeepSeek swapped for 53 hours

Separately, Prime Intellect disclosed on November 19 (advisory PI-SA-2026-001) that an unauthorized party replaced the model behind its deepseek/deepseek-v4.1-flash endpoint with an "abliterated" copy — public weights modified to reduce or remove refusals. The swap lasted from November 8 (01:24 UTC) to November 10 (06:35 UTC), 53 hours and 11 minutes. The company's tests show capability did not change (GPQA Diamond and τ-bench stable), but refusals collapsed: from 98 out of 100 prompts to 1 on JailbreakBench, from 304 out of 313 to 2 on StrongREJECT, and from 198 out of 240 to none on HarmBench. Prime Intellect says it found no evidence that prompts or completions were recorded or left the cluster — but OpenRouter's email to affected customers, shared on X by Theo, the developer behind T3 Chat, said it could not be ruled out that requests had been read while being processed. The company says it contacted all affected customers and recommends reviewing outputs from the period; the investigation continues.

The lesson of the prompt theft is blunt: a privacy flag binds only the first hop — agents send whatever they can read, credentials included. The advice is to pick upstream providers by hand instead of accepting the default pool, verify that privacy settings actually change where requests go, and, for those who cannot take the risk, host the model. A test with fake credentials planted in the context works as a tripwire: any use of them should be treated as an incident.

Image: OpenRouter illustration with the "Zero Data Retention" seal crossed out, by The Hacker News.

Sources: The Hacker News; Prime Intellect — Notice regarding deepseek/deepseek-v4.1-flash (PI-SA-2026-001); CryptoNews; CybersecurityNews.

Transparency note: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet, and reviewed before publication. Check the original sources to confirm the information.

Por Muse