Tag: Phala

  • OpenRouter Off-Boards Providers After Data-Processing Incidents (2026)

    OpenRouter Off-Boards Providers After Data-Processing Incidents (2026)

    Takeaway: On November 22, 2026, OpenRouter published Data Processing Incidents on Inference Providers. The company says security researchers reported that certain inference providers had data compromised or otherwise failed to uphold their data policies. OpenRouter states that the findings did not show any breach or unauthorized access of OpenRouter itself. It says it off-boarded three providers, removed 125 endpoints from Zero Data Retention (ZDR) eligibility, and notified all potentially affected customers.

    The 125 removals, OpenRouter writes, included providers flagged by the researchers plus additional endpoints taken down as a precaution. The post thanks the Dragos canary-token study and says OpenRouter is adapting similar tests for ongoing checks. This article follows that official post and the three provider write-ups OpenRouter listed. It does not treat those three public posts as a list of the off-boarded providers, because OpenRouter did not name the three providers it removed.

    What OpenRouter says it changed

    OpenRouter describes itself as the operating layer between customers and a mix of models and providers, and says it takes data-policy enforcement seriously. Besides off-boarding three providers, removing 125 ZDR endpoints, and notifying potentially affected customers, it lists work still underway:

    • Adapting the researchers’ testing approach so OpenRouter can run those checks on an ongoing basis.
    • Technical verification of network paths to identify cases where requests and responses may be forwarded to third parties.
    • Requiring appropriate assurances and validation before restoring ZDR eligibility to any endpoint removed in this review.

    The company also says it is encouraging all providers to publish incident write-ups and will keep its list updated. “Not all affected providers have posted publicly,” the announcement adds. “We have notified affected users regardless of whether there is a public post.”

    Provider posts OpenRouter listed

    OpenRouter’s live list, as of the November 22 post, pointed to three public notices: Novita’s privacy notice about MiniMax requests, Phala’s November preliminary incident report, and Together’s security event about prompt leakage for some June and July 2026 requests. Those are the only provider statements this article summarizes. OpenRouter did not say these three accounts are the only affected providers, or that they are the three it off-boarded.

    Novita: MiniMax requests forwarded off-platform

    In a privacy notice dated November 22, 2026, Novita AI says an independent researcher reported a problem in how MiniMax requests were handled on its platform. Novita’s review confirmed that some requests were forwarded to MiniMax’s official API rather than processed on Novita-hosted infrastructure. It says a routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly, and that it promptly suspended the affected models.

    If you used MiniMax models through Novita between May 28, 2026, and November 11, 2026, Novita says your request contents may have been accessed or retained by a third party. That window includes requests made directly through Novita and those routed to Novita through OpenRouter. The company says the researcher observed accesses to URLs and hostnames embedded in test requests, and that no use of the AWS test key markers was observed. It describes the access patterns as suggesting that automated content review may have been involved, and says it is following up with MiniMax to verify that assessment and the handling of any affected data.

    Novita states that it does not retain customer prompts or generated outputs or use them for model training. It is verifying processing paths for the affected MiniMax endpoints and asking MiniMax about any retained customer data and its deletion. If requests included live credentials, API keys, or other secrets, Novita recommends revoking or rotating them as a precaution. Questions can go to privacy@novita.ai.

    Phala: DeepSeek V4 Flash 0731 requests exposed

    Phala’s November preliminary incident report, dated November 10, 2026, says two groups of users may need to take action. First, users whose DeepSeek V4 Flash 0731 requests were routed to Phala by OpenRouter between October 21, 2026, at 00:00 UTC and November 8, 2026, at 05:35 UTC should treat request contents as exposed and rotate any API keys, tokens, passwords, or other secrets included in those requests. Second, owners of affected confidential virtual machines (CVMs) were contacted directly: Phala says the attacker obtained application credentials, and affected CVMs may also have had secrets derived from Phala’s key management system exposed.

    Phala says it contained the incident on November 8, restored the affected model service with additional protections and monitoring, and has observed no successful unauthorized access since containment, including after further attacker attempts that followed a KMS root-key rotation on November 6. The company is still investigating and says it is limiting exploit-specific details until a full root-cause analysis is ready. This article does not add those details.

    At a high level, Phala reports unauthorized activity first detected on November 5, involving a legacy internet-facing service and management interfaces that trusted loopback access rather than independently authenticating callers. It says the attacker created unauthorized CVMs and obtained application credentials. On November 8, a third-party researcher provided evidence that request data for DeepSeek V4 Flash 0731 had been exposed. Phala traced that evidence to an inference deployment on a new GPU server that was serving production traffic without its full production security hardening, and says in-flight request data was visible during processing. It took the deployment and model offline the same day.

    The report’s precautionary exposure window for that model starts on October 21. Phala also notes that a researcher regularly included unique canary credentials in requests from July 1 through November 7, and that only canaries from requests sent on October 21 and 22 were later misused. Because those tests did not cover every request, Phala is keeping the wider window for user guidance. It says a full-platform audit found no evidence of compromise affecting other models or products, and that Phala served this model only through the affected route.

    Together AI: stored prompt samples from June and July

    Together AI’s CEO and co-founder, Vipul Ved Prakash, published a security event notice dated November 21, 2026, about a subset of prompts submitted to Together’s serverless inference service in June and July 2026. Together says an unauthorized party downloaded stored prompt samples after a credential for a private repository was exposed in a public code repository.

    According to the post, Together collected prompt samples authorized for routine software testing and stored them in a private Hugging Face repository. The samples came from requests to MiniMax M3, GLM-5.2, and Kimi K2.6, primarily on June 24, July 4, July 8, and July 14, 2026. During development of Together’s public GitHub repository, xorl, an AI coding agent committed a credential that provided access to that private repository. Together says automated scanning detected the exposed credential and the commit was reverted. During the exposure window, an unauthorized party obtained the credential, downloaded the stored samples, searched them for credentials to other services, and tested whether those credentials were valid. Five of the credentials tested, Together writes, were canary credentials associated with a Dragos security experiment.

    Together says it has worked with an external forensic investigation firm and law enforcement. Based on its review, it has not identified evidence of additional prompt leakage beyond this incident. A review of the affected prompt contents is still underway to determine whether sensitive personal information or other customer data was exposed, and Together says it will contact affected customers as needed. After the incident, it says it strengthened controls around development workflows, credentials, and customer data, including scoped or short-lived credentials and keyless access where supported, credential checks on all pull requests, making ZDR the default setting for new accounts, and updating development processes for AI coding agents.

    Customers handling sensitive prompt data, Together writes, should enable ZDR in their account settings. Questions can go to security@together.ai.

    How this fits earlier coverage

    These notices sit on top of Dragos’s November 19, 2026 canary study, which we covered in Dragos Canary Study Finds AI Prompt Leaks and Provider Breaches. That post reported evidence of breaches at three unnamed providers, reseller logging, and cases where privacy or ZDR settings were not honored. OpenRouter’s November 22 announcement is the company’s own account of what it changed after that research, not a second telling of the Dragos findings. OpenRouter also remains the routing layer behind a recent Hugging Face change; see Hugging Face and OpenRouter Inference Routing: What Changed. Together’s description of a credential committed by an AI coding agent is a separate incident from the earlier GitHub secret-leak pattern we described in Claude Code settings.local.json Secrets on GitHub.

    Key facts

    ItemWhat official sources say
    OpenRouter announcementNovember 22, 2026. Findings did not show a breach of OpenRouter itself.
    OpenRouter actionsOff-boarded three unnamed providers; removed 125 endpoints from ZDR eligibility, including researcher-flagged and precautionary removals; notified all potentially affected customers.
    Still underwayOngoing versions of the researchers’ tests; network-path verification for third-party forwarding; assurances required before any removed endpoint regains ZDR eligibility.
    NovitaSome MiniMax requests from May 28 to November 11, 2026, including OpenRouter-routed traffic, were forwarded to MiniMax’s official API. Rotate secrets if they were in those requests.
    PhalaDeepSeek V4 Flash 0731 requests routed by OpenRouter from October 21, 2026, 00:00 UTC, to November 8, 2026, 05:35 UTC, should be treated as exposed. Incident contained November 8. Other models: no evidence of compromise, according to Phala.
    Together AIStored prompt samples from MiniMax M3, GLM-5.2, and Kimi K2.6, primarily on June 24 and July 4, 8, and 14, 2026, were downloaded after a private-repository credential was committed to a public GitHub repo. Together reports no evidence of additional prompt leakage beyond this incident and is still reviewing contents.
    Unnamed providersOpenRouter says not every affected provider has posted publicly, and that it notified affected users anyway.

    FAQ

    Was OpenRouter itself breached?

    OpenRouter says no. Its November 22 post states that the researchers’ findings did not show any breach or unauthorized access of OpenRouter. The issues it describes concern inference providers in its ecosystem, and the actions it lists are off-boarding, ZDR-eligibility changes, and customer notification.

    Which three providers did OpenRouter off-board?

    OpenRouter did not name them. It listed three public provider write-ups — from Novita, Phala, and Together — and separately said not all affected providers have posted publicly. This article does not treat that public list as the off-boarding list.

    What does removing 125 ZDR endpoints mean?

    OpenRouter says it removed 125 endpoints from Zero Data Retention eligibility. The removals included providers flagged by researchers and additional endpoints removed as a precaution. It is requiring assurances and validation before any of those endpoints can regain ZDR eligibility.

    Which Novita requests are in scope?

    Novita’s November 22 notice covers MiniMax model requests made through Novita between May 28, 2026, and November 11, 2026, including requests routed to Novita through OpenRouter. Novita says some of those requests were forwarded to MiniMax’s official API, and it recommends rotating any live credentials included in them.

    What should OpenRouter users of Phala’s DeepSeek model do?

    Phala says users whose DeepSeek V4 Flash 0731 requests were routed to it by OpenRouter between October 21, 2026, 00:00 UTC, and November 8, 2026, 05:35 UTC, should treat request contents as exposed and rotate API keys, tokens, passwords, or other secrets included in those requests. Phala also says it contained the incident on November 8 and found no evidence that other models or products were compromised.

    What did Together AI disclose?

    Together says a subset of serverless prompt samples from MiniMax M3, GLM-5.2, and Kimi K2.6, primarily on four dates in June and July 2026, were stored in a private Hugging Face repository and later downloaded after an AI coding agent committed a repository credential to Together’s public xorl GitHub repo. Together reports no evidence of additional prompt leakage beyond this incident and says it will contact affected customers as needed.

    Did OpenRouter notify customers?

    Yes. The company says it notified all potentially affected customers, including in cases where a provider has not published a public incident post.

    Is this the same story as the Dragos canary study?

    It is a follow-on, not a substitute. Dragos published its canary-token findings on November 19, 2026, without naming the three providers it said were compromised. OpenRouter’s November 22 post is the router’s account of actions it took after that research, plus a list of provider notices it had received. For the study itself, see our earlier article on the Dragos report.