OpenRouter announced this Sunday (22) the removal of 125 endpoints from the Zero Data Retention (ZDR) category and the shutdown of three inference providers after an investigation identified flaws in the protection of data sent to artificial intelligence models. The problems include exposure of prompts, improper access and forwarding of information to third parties without the privacy guarantees expected by clients.

The company said the endpoints lost eligibility for the zero-retention badge, including services removed preventively. All potentially affected clients were notified. OpenRouter stressed that there was no evidence of intrusion or unauthorized access to its own platform.

The decision came after research published on November 19 by cybersecurity firm Dragos. Led by researcher HD Moore, the study sent approximately 2 million test prompts to AI services between July and November 2026, using unique URLs and monitored credentials to identify possible exposures.

Researchers observed later accesses to the addresses inserted in the requests, including by organizations that should not receive that data. In some cases, credentials used as bait were tested by third parties. The investigation also identified requests sent to US providers that ended up reaching Chinese organizations.

In one of the cases involving MiniMax models, a notice sent by OpenRouter to clients, reproduced on Hacker News, reported accesses to the test URLs between 40 and 130 hours after the requests. The records indicated that the content could remain accessible after the response was processed.

What the zero-retention badge guaranteed

The Zero Data Retention classification indicates that prompts and responses should not be stored by the service responsible for inference. The problem identified by Dragos is in the processing chain: a request may pass through different intermediaries before reaching the infrastructure that runs the model.

Each intermediary may have access to the full content of the request. Thus, the retention policy declared by the first service does not by itself guarantee that all subsequent providers follow the same rule.

Novita confirmed, in a statement published on November 22, that a routing configuration caused requests destined for MiniMax models to be forwarded to the developer's official API, instead of remaining on Novita's infrastructure. The company warned that requests made between May 28 and November 11, 2026 may have been accessed or stored by third parties.

In a communication reproduced publicly, OpenRouter also reported that MiniMax had declared to StreamLake that its content moderation mechanisms were active before October 22 and that, since then, it had begun enabling ZDR. Until that moment, however, the developer had not presented directly to OpenRouter its own explanation of the episode.

Other providers disclosed separate incidents. Together AI acknowledged that samples of prompts from June and July were obtained by an unauthorized person after the exposure of an access credential to a private repository. Phala reported the compromise of infrastructure used to process requests to the DeepSeek V4 Flash 0731 model.

OpenRouter cited Novita, Phala and Together among the companies with public statements, but did not name the three providers that were shut down. Nor did it state that all 125 removed endpoints suffered proven leaks.

In response, the platform announced that it intends to incorporate tests similar to those used by Dragos, technically verify the paths taken by requests and require additional validations before restoring the ZDR classification of the affected endpoints.

For companies that use AI APIs with sensitive data, the episodes highlight the importance of verifying not only the contracted privacy policy, but also the identity of the effective providers, possible traffic forwarding and storage conditions. Dragos recommends restricting intermediaries, using bait credentials to detect exposures and avoiding the unnecessary sending of secrets to models.

Tools mentioned

More from Radar