Three inference providers disclosed prompt failures in the same week
Together AI, Novita, and OpenRouter each disclosed a different way prompts were mishandled — stored, misrouted, or passed on. The failures are in the plumbing, not the models.
- Together AI says an unauthorized party downloaded stored prompt samples after a credential was exposed in a public code repository; the samples came from requests to MiniMax M3, GLM-5.2, and Kimi K2.6 in June and July 2026. Self-reported m:…9271
- Together AI says an AI coding agent committed the credential during development of its public GitHub repository, xorl; its automated scanning detected the exposure and the commit was reverted. Self-reported m:…9271
- Novita says some MiniMax requests were forwarded to MiniMax's official API instead of Novita-hosted infrastructure between May 28 and November 11, 2026, covering requests made directly and through OpenRouter; Novita suspended the affected models. Self-reported m:…0506
- OpenRouter says it off-boarded three providers and removed 125 endpoints from zero data retention eligibility after researchers reported data policy violations, and says there was no breach of its own platform. Self-reported m:…0062
- Together AI says it is making zero data retention the default for new accounts, and Novita recommends revoking or rotating any live credentials, API keys, or other secrets included in requests during the affected period. Self-reported m:…9271 m:…0506
Three disclosures, three different failures
Three hosted inference providers disclosed prompt-handling failures in the same week, and each one failed in a different place. Together AI says an unauthorized party downloaded stored prompt samples after a credential was exposed in a public code repository. Novita says a routing error sent some MiniMax requests to MiniMax's official API instead of Novita's own infrastructure. OpenRouter says it off-boarded three providers and removed 125 endpoints from zero data retention eligibility after researchers reported data policy violations. m:…9271 m:…0506 m:…0062
None of the three is an attack on a model. Each is a provider-side failure in how prompts were stored, routed, or passed on — and each raises the same question for developers: when you send a prompt to a hosted provider, where does it actually go? m:…9271 m:…0506 m:…0062
Together: a credential committed by a coding agent
Together AI's incident reaches back to June and July 2026. The company says a subset of prompts sent to its serverless inference service was affected, with samples drawn mainly from requests on June 24, July 4, July 8, and July 14. The samples came from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, and were stored in a private Hugging Face repository. m:…9271
The credential that opened that repository was committed by an AI coding agent during development of Together's public GitHub repository, xorl. Together says its automated scanning detected the exposed credential and the commit was reverted, but an unauthorized party had already downloaded stored prompt samples. m:…9271
The company says it has worked with an external forensic investigation firm and law enforcement, and says it has no reason to believe the data was leaked further and understands it has now been deleted. Its review of whether sensitive personal information or other customer data was exposed is still ongoing. m:…9271
Novita: requests sent to the wrong API
Novita's failure was routing, not a leak of stored data. The company says it received a report from an independent security researcher about how MiniMax requests were handled, and its review confirmed that some requests were forwarded to MiniMax's official API rather than processed on Novita-hosted infrastructure. A forwarding path meant for specific customers had been applied more broadly. m:…0506
The affected window runs from May 28 to November 11, 2026, and covers requests made directly through Novita and requests routed to Novita through OpenRouter. The researcher observed accesses to URLs and hostnames embedded in test requests, which suggests the forwarded traffic was read somewhere along the way. m:…0506
Novita says it suspended the affected models, is verifying the processing paths, and is following up with MiniMax about retained data and its deletion. It recommends that anyone whose requests included live credentials, API keys, or other secrets revoke or rotate them. Novita says it does not itself retain customer prompts or generated outputs or use them for model training. m:…0506
OpenRouter: three providers off-boarded, 125 endpoints cut
OpenRouter sits in the middle of this picture: it routes requests to providers, and it says researchers reported earlier in the week that certain providers had data compromised or had shown evidence that their data policies were not being upheld. OpenRouter says it off-boarded three providers and removed 125 endpoints from zero data retention eligibility, and notified all potentially affected customers. It says there was no breach or unauthorized access of its own platform. m:…0062
The company says it is adapting the researchers' testing approach so it can run the checks on an ongoing basis, and is verifying network paths to find cases where requests and responses may be forwarded to third parties. It says it will require assurances and validation before restoring zero data retention eligibility to any endpoint removed in the review. m:…0062
What to check before you trust a provider
The three disclosures point to the same checks. First, whether zero data retention is actually on: Together says it is making it the default for new accounts, and OpenRouter's removals show that eligibility is not a guarantee — 125 endpoints lost it in one review. Second, where requests are processed: Novita's forwarding path shows that a request sent to one provider can end up at another company's API. Third, what was in the prompts: Novita's advice to rotate any credentials sent during the affected window is the practical response when you cannot be sure what was read. m:…9271 m:…0062 m:…0506
The disclosures do not say how many prompts were affected in any of the three cases, and Together's review of what was in the exposed samples is not finished. What they do say is that the trust question is not about the model — it is about the plumbing between you and it. m:…9271 m:…0506
What is still unresolved
- Whether sensitive personal information or other customer data was in the prompt samples Together AI says were downloaded — the company's review is ongoing.
- Whether MiniMax retained request data forwarded by Novita, and whether it has been deleted — Novita says it is still following up with MiniMax.
- Which three providers OpenRouter off-boarded — they are not named in the disclosure.
- How many prompts were affected in each of the three cases — none of the disclosures gives a number.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.