OpenRouter cuts 3 AI providers over leaked customer prompts
Part of the OpenRouter story · since 22 Nov
- Sun 22 Nov, 02:55 UTC. cybersixt, first report. OpenRouter cuts 3 AI providers over leaked customer prompts. This article, 1st of 1.
OpenRouter, an AI model router and marketplace, has off-boarded three inference providers and removed 125 endpoints from its Zero Data Retention (ZDR) list after security researchers showed that customer prompts were leaking. In a blog post on Sunday, OpenRouter says the findings showed no breach of OpenRouter itself, and that it has notified all potentially affected customers. It did not name the three providers it dropped.
The research came from Dragos, which said on Thursday that it had sent about two million prompts containing unique URLs and canary credentials to AI services from July to November and watched where they surfaced. Dragos says data from at least three providers was exposed, that resellers on an inference marketplace could read and keep customer prompts with ordinary logging, and that some providers did not honour privacy settings. Two providers have now said what went wrong on their side, alongside a third report from Phala.
Together AI: a key committed by a coding agent
Together AI says an unauthorised party downloaded stored prompt samples from its serverless inference service. The samples, kept in a private Hugging Face repository for software testing, came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, mainly on June 24, July 4, July 8 and July 14. While developing its public GitHub repository xorl, an AI coding agent committed a credential for that private repository. Together's scanning caught it and the commit was reverted, but in the meantime someone used the key, downloaded the samples, searched them for credentials to other services and tested them. Five of those were Dragos canaries.
Together says it has worked with a forensic firm and law enforcement, has found no evidence of other prompt leakage and is still reviewing what personal data the samples held. It has made ZDR the default for new accounts, moved to scoped or short-lived credentials, checks every pull request for secrets and has changed how it uses AI coding agents.
Novita: requests forwarded to MiniMax
Novita says a researcher showed that some MiniMax requests were forwarded to MiniMax's own API instead of being processed on Novita's infrastructure, because a forwarding path meant for specific customers applied more broadly. Anyone who used MiniMax models through Novita between May 28 and November 11, directly or routed through OpenRouter, may have had request contents accessed or retained by a third party. Novita says the researcher saw URLs embedded in test requests being visited, though no use of the AWS test keys, and that the pattern suggests automated content review. It has suspended the affected models and is following up with MiniMax about retained data.
Phala: an inference server taken over
Phala published a preliminary report on November 10. An attacker chained a legacy internet-facing service with management interfaces that trusted local connections, created confidential virtual machines and took application credentials, then took control of an inference server that was serving traffic without full hardening. Users whose DeepSeek V4 Flash 0731 requests were routed to Phala by OpenRouter between October 21 and 05:35 UTC on November 8 should treat those requests as exposed, Phala says.
Separately, Prime Intellect disclosed this month that an attacker had swapped its DeepSeek deployment for a modified copy.
What to do
Novita and Phala both tell affected users to rotate any API keys, tokens, passwords or other secrets included in their requests. Together advises customers handling sensitive prompts to turn on ZDR. OpenRouter says it is building the researchers' tests into ongoing checks, verifying network paths to find requests forwarded to third parties, and will require assurances before any removed endpoint regains ZDR status.
OpenRouter says not every affected provider has posted publicly and that it will update its list as more write-ups arrive.
How the story developed
1 article from 1 outlet since 22 Nov. Still developing; the latest report was 52 minutes ago. This was the first report.
- Sun 22 Nov, 02:55 UTC. cybersixt, first report. OpenRouter cuts 3 AI providers over leaked customer prompts. This article, 1st of 1.
- No CVE
- Larger: an outlet’s first report
- Coverage over time
Show all 1 articles as a listHide the list
- Sun 22 NovDay 01
- OpenRouter cuts 3 AI providers over leaked customer prompts02:55 · cybersixt · First report · This article, 1st of 1
Sources
- OpenRouter: Data processing incidents on inference providersPrimary sourceopenrouter.ai
- Dragos: AI supply chain security, what canary tokens foundPrimary sourcedragos.com
- Together AI: Security event, prompt leakage for requests in June and July 2026Primary sourcetogether.ai
- Novita: Privacy notice regarding MiniMax requestsPrimary sourcenovita.ai
- Phala: November preliminary incident reportPrimary sourcephala.com
- CyberSIXT: Prime Intellect DeepSeek model swapcybersixt.com
More reporting
- Fluidkey hack: $1.6M stolen via session tokens in SigNoz logs22 Nov
- GoFundMe data breach: nonprofit donor data accessed via API21 Nov
- iTerm2 vulnerability: cat on a crafted file runs commands21 Nov
- Upload2Shell: Ghostscript CVE-2026-125396 lets uploads run code21 Nov
- GL.iNet GoodCloud flaw used to plant malware on routers, KVMs21 Nov
Written by CyberSIXT from the sources above.How we reportCorrectionsRSS