Skip to content
[ SIGNAL VERIFIED ] INDEPENDENT CYBERCRIME JOURNALISMCLASSIFIED // THE UNDERGROUND HAS A NEWSROOM / RSS
.COM
Latest News
ALL NEWS
Newsroom

OpenRouter Removes Three AI Providers as Prompt Privacy Failures Strip 125 Endpoints of ZDR Status

OpenRouter removed three providers and withdrew Zero Data Retention eligibility from 125 endpoints after provider-level prompt leaks and routing failures came under scrutiny.

Newsroom editorial artwork
PUBLISHED

IPCRIME | November 22, 2026 — evening newsroom update. AI model gateway OpenRouter says it has removed three inference providers and withdrawn Zero Data Retention (ZDR) eligibility from 125 endpoints following researchers’ findings about prompt exposure, provider compromise and data-processing practices. The gateway says the research did not identify a breach or unauthorized access of OpenRouter itself. The problems were observed in downstream hosting and routing relationships.

What OpenRouter changed

In its November 22 announcement, OpenRouter says it removed three providers, applied both targeted and precautionary ZDR eligibility restrictions, notified potentially affected users and began independently verifying where prompts travel. The 125 endpoints are not 125 confirmed data breaches. The company has not named the three providers it off-boarded, and its list of provider notices must not be assumed to name those three.

What the researchers actually tested

Dragos researchers submitted approximately two million synthetic prompts seeded with unique URLs and credentials to providers between July and November. Their monitors subsequently observed some of those planted secrets used in attempts to access controlled test services. The tests exposed multiple kinds of failure: unauthorized access at some hosting companies, routine marketplace-provider visibility into prompts, and routes or retention practices inconsistent with expectations. This method does not establish a global number of exposed real customers.

Together AI: a repository credential exposed stored prompts

Together AI reported on November 21 that an AI coding agent inadvertently committed a credential to a public code repository. Someone used it to download samples of customer prompts from a private Hugging Face repository, then searched some samples for other credentials and tested them. The records mainly concerned MiniMax M3, GLM 5.2 and Kimi 2.6 requests in June and July. Together is still reviewing the scope and said it had no reason to believe the copied material was further distributed.

Phala: a compromised inference environment

Phala’s November 10 preliminary report acknowledges that unauthorized access to an inference environment made some DeepSeek V4 Flash 0731 prompt contents visible. Its precautionary window covers requests routed to Phala between October 21 at 00:00 UTC and November 8 at 05:35 UTC, though observed use of planted canaries was narrower. Phala also reported application-credential exposure affecting confidential virtual-machine infrastructure. It says the recognized attack paths were contained and advises potentially affected customers to rotate secrets included in requests.

Novita: an unintended third-party forwarding route

Novita’s November 22 privacy notice says a routing configuration mistake sent some MiniMax requests to MiniMax’s official API rather than handling them fully on Novita-hosted machines. Requests through Novita between May 28 and November 11 may have been processed or retained by that third party. Novita observed that unique test URLs were accessed but says planted AWS test credentials were not used. This is evidence of a routing and data-handling failure, not proof that all forwarded prompts were maliciously stolen.

Why a ZDR label is not an encryption guarantee

Zero Data Retention is a processing promise, not end-to-end encryption of the text between the author and the inference engine. An AI hosting service generally needs to read the prompt to process it. If a routing intermediary forwards the request outside the expected chain, or a provider is breached while handling it, the privacy label alone does not prevent exposure. Service boundaries, provider selection and credential handling therefore matter as much as account settings.

Practical response for customers

  • Identify which models, provider routes and time periods handled sensitive data. Check your provider’s account notices and actual routing logs rather than assuming every AI request was exposed.
  • Revoke and replace any real API keys, passwords or tokens included in affected prompts; review corresponding authentication and cloud activity.
  • Audit applications that sent private client information or unpublished source code to inference platforms. Limit future submissions to the minimum necessary data.
  • Review retention agreements, sub-processors, usage policies and data residency for each stage of the inference chain.
  • Use narrowly scoped, short-lived test credentials and synthetic values when developing applications.

What remains unresolved

OpenRouter has not identified all three removed providers by name. Dragos’s account of at least three compromised providers should not be mapped automatically onto OpenRouter’s separate list of provider actions. Nor has a final count of affected production prompts been established. The responsible response is to follow the distinct notices and update risk assessments as facts emerge.

Original sources and methodology

OpenRouter — November 22 provider response; Dragos — canary study; Together AI — prompt-leak report; Phala — preliminary incident report; Novita — forwarding notice. IPCRIME independently synthesized public records. We did not test provider endpoints, access customer prompts or make a claim of an OpenRouter system breach.

Editorial note

IPCRIME distinguishes public records and documented findings from allegations, analysis and commentary. Report corrections through our contact page.