Subscribe
Cybersecurity

OpenRouter drops three AI hosting providers after months of prompt theft

Sunday, November 22, 2026 · 6 min read

OpenRouter removed three AI hosting providers and stripped 125 endpoints of Zero Data Retention eligibility after Dragos seeded about two million prompts with canary tokens and caught months of prompt theft — including a Phala window from Oct 21 to Nov 8 and a Together AI credential leak.

Bearish$CRWD$PANW

📊 Scorecard: our Bearish call on $CRWD is scored from $229.44, the last close before it ran. Its first result comes after the next close.

OpenRouter drops three AI hosting providers after months of prompt theft

The News

OpenRouter said on Sunday, Nov 22, that it has removed three AI hosting providers, removed 125 endpoints from Zero Data Retention eligibility, and notified customers who may have been affected. The move follows canary-token research by Dragos, an industrial cybersecurity company, which seeded about two million prompts with unique links and credentials between July and November. Dragos said attackers stole data inside three AI inference providers for months, and none of the three knew until the researchers told them. OpenRouter said the findings showed no breach of OpenRouter itself.

The largest breach began with a couple hundred planted AWS keys, each sent in exactly one request to one provider in a few short windows in late October. Every request sent during those windows was stolen — including some the provider had rejected before a model ever ran — and the thief then drove tens of thousands of AWS API calls, listing S3 buckets, searching Secrets Manager, and trying to create Bedrock API keys and add IAM users. Some stolen requests had been flagged Zero Data Retention. Dragos assesses real customer requests were almost certainly copied too — an assessment, not a proven fact.

The providers have started naming what happened. Phala said an attacker took over an inference server and could view request data while it was being processed; requests to its DeepSeek V4 Flash 0731 model via OpenRouter between Oct 21 at 00:00 UTC and Nov 8 at 05:35 UTC should be treated as exposed, with every key, token or password in them rotated. The incident was contained on Nov 8. Together AI said an unauthorised party downloaded stored prompt samples after an AI coding agent committed a credential for a private Hugging Face repository to a public GitHub project; the samples came from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, mainly on June 24 and July 4, 8 and 14. Novita said a routing issue sent some MiniMax requests on to MiniMax's official API; requests between May 28 and Nov 11 may have been accessed or retained by a third party.

A second finding needed no break-in. Sellers on an AI inference marketplace listed models at the lowest prices, switched on request logging and collected what flowed through — one operator in France took passwords from two prompts and typed root SSH login attempts by hand. An unnamed Chinese AI 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 of those marked no data collection or Zero Data Retention. Neither Dragos nor OpenRouter has named its three, and neither has said whether they are the same three.

The cheapest AI hosting can carry the highest price — your prompts, your keys, your data, read by strangers.

Why It Matters

This matters because prompts carry the secrets of modern software. Developers paste API keys, passwords and private documents into AI requests every day, and a gateway hands those prompts to the provider running the model. If that provider is compromised, every secret in every prompt is exposed — and the thief here did not stop at reading prompts: tens of thousands of AWS API calls show what stolen keys are worth.

The Zero Data Retention failure should worry enterprises most. Customers pay for that setting because they do not trust providers to store their data — yet flagged requests were stolen anyway, and an unnamed Chinese lab fetched the secret links from about two-thirds of requests marked no data collection or Zero Data Retention. A privacy promise that fails silently is worse than no promise at all.

The market angle runs through the read-throughs. CrowdStrike and Palo Alto Networks are the public security vendors closest to this story, but the breached companies are private — OpenRouter, Dragos, Phala, Together AI and Novita have no ticker — so $CRWD and $PANW are read-throughs, not the breached party. The breach is bad for the AI-hosting vendors, not the security vendors, which is why both tickers carry a mixed call against a bearish story verdict.

Analyst's Take

Our call is Bearish. Months of prompt theft at three AI hosting providers, a Zero Data Retention promise that did not hold, and planted keys turned into tens of thousands of cloud API calls — a trust problem for the whole AI-hosting layer. The breach is bad for the AI-hosting vendors, not for the security vendors — which is exactly why this post carries mixed calls on CrowdStrike and Palo Alto Networks rather than bearish ones.

What to watch next is whether OpenRouter names the three providers it removed, and whether they match the three Dragos found breached — neither side has said. Also watch whether any affected customer reports real damage, and whether other gateways audit their providers. The clearest action item is the Phala window: anyone who sent keys, tokens or passwords to DeepSeek V4 Flash 0731 via OpenRouter between Oct 21 and Nov 8 should rotate them.

For investors, the read-through is that security spending follows trust failures. CrowdStrike and Palo Alto Networks sell monitoring and credential protection that catches exactly this sort of theft, which is why their call here is mixed rather than bearish — a demand signal for their products, even as it drags the sector's mood. The deeper question is whether cheap inference stays cheap once customers ask what their prompts are worth. Reader poll: Would you send secrets through an AI gateway — yes or no? Vote: Yes, I trust it, No, never, Only with ZDR.

Are you bullish or bearish on this story?

By The Numbers

~2MPrompts Dragos seeded with canary tokens, July–November
3Providers removed by OpenRouter; three breached, per Dragos
125Endpoints removed from Zero Data Retention eligibility
Oct 21–Nov 8Phala DeepSeek V4 Flash exposure window via OpenRouter
Tens of thousandsAWS API calls driven by the thief after the largest breach
>50%Seeded prompts whose secret links an unnamed Chinese lab fetched

FAQ

What is Zero Data Retention?

Zero Data Retention is a setting that is supposed to mean a prompt sent to an AI provider is never stored — it is processed and then forgotten. In this story, some requests flagged Zero Data Retention were still stolen, which shows the promise depends on the provider actually honouring it, not just offering the checkbox.

What is a canary token?

A canary token is a planted trap: a unique web link, password or API key that exists only inside one test prompt. If anyone ever uses it, the researchers know exactly which prompt was stolen, from which provider, and when. Dragos seeded about two million prompts with canary tokens between July and November.

What is an AI inference provider?

An inference provider is a company that runs AI models on its own servers and answers requests from apps and users. OpenRouter is a gateway that routes customer requests to many such providers. When you send a prompt through a gateway, the provider that actually runs the model sees everything in it.

What is a prompt, and why does stealing one matter?

A prompt is the text a user sends to an AI model — a question, a document, or a chunk of code. Prompts often contain secrets such as passwords, API keys and private documents. Stealing prompts can hand an attacker the keys to a company's cloud accounts.

👁

Source: The Hacker News; Together AI and Novita security notices (sample dates and exposure window)

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $CRWD

All $CRWD coverage →

More from AI Agent Update