OpenRouter Boots Three Providers, Strips 125 Endpoints of Zero-Retention Status
OpenRouter off-boarded three inference providers and pulled 125 endpoints from Zero Data Retention after Dragos canary tests exposed prompt leakage across routing chains.
By Grace Kim
4 min read

What's News
- OpenRouter off-boarded three providers and removed 125 endpoints from ZDR eligibility on November 22, 2026.
- Dragos submitted roughly 2 million canary prompts from July to November 2026 and traced about 2,000 inference routes.
- Seeded AWS keys were used for tens of thousands of API calls after exposure in late October 2026.
- Phala's breach window ran from October 21 to November 8, 2026; Together AI's leak stemmed from a committed Hugging Face credential.
- One unnamed Chinese AI lab fetched secret URLs from over half of prompts, including two-thirds marked ZDR.
OpenRouter removed three inference providers and pulled 125 endpoints from Zero Data Retention (ZDR) eligibility on November 22, 2026, after researchers demonstrated that customer prompts were being exposed across the routing chain. The company also notified all potentially affected customers. It stressed the findings showed no breach of OpenRouter itself.
The public list links incident notices from Novita, Phala and Together AI, though OpenRouter doesn't name them as the three off-boarded providers. Not every affected provider has published a notice. The removals mix flagged providers with endpoints withdrawn purely as a precaution.
What did the canary tests actually catch?
On November 19, 2026, Dragos published research describing roughly two million prompts containing unique URLs and canary credentials, submitted to widely used AI services between July and November 2026. The monitoring that followed produced three hard findings:
- Data from at least three AI providers was exposed.
- Marketplace resellers viewed and retained customer prompts through standard logging.
- Some providers appeared to disregard privacy settings or their own published policies.
The largest compromise involved a couple hundred seeded AWS keys, sent during several short request windows in late October 2026. Some keys came from requests rejected before a model even ran — Dragos read that as evidence the copying happened when requests first reached the provider, not during processing. The bait keys were then used for tens of thousands of AWS API calls.
The routing picture matters because every organization on a multi-hop path can see the full prompt and response. Dragos traced about 2,000 inference routes. Several hundred included an unidentified system; in a handful, a company absent from the advertised path fetched the test image. One unnamed Chinese AI lab's systems fetched secret URLs from more than half the prompts sent to its models between July and October 2026 — including roughly two-thirds marked no-data-collection or ZDR.
How did each provider get breached?
Novita's routing configuration broadened an API forwarding path meant for specific customers, sending some MiniMax requests to MiniMax's official API instead of Novita-hosted infrastructure. Requests made directly through Novita, or routed through OpenRouter between May 28, 2026, and November 11, 2026, may have had contents accessed or retained by a third party. Novita suspended the affected models. The researcher observed accesses to test URLs and hostnames but no use of AWS test-key markers; automated content review may have been involved, an assessment Novita is checking with MiniMax.
Phala's preliminary report, published November 10, 2026, told users whose DeepSeek V4 Flash 0731 requests were routed by OpenRouter between 00:00 UTC on October 21, 2026, and 05:35 UTC on November 8, 2026, to treat contents as exposed and rotate any included secrets. The attacker chained a legacy internet-facing service with management interfaces that trusted loopback access without authenticating callers. A similar pattern gave the attacker control of an inference VM, making in-flight request data visible during processing. Phala confirmed canary misuse only on October 21–22, 2026, but kept the wider window because sampling can't reveal every access event. Containment finished November 8, 2026; a full-platform audit found no compromise of other models or products.
Together AI disclosed on November 21, 2026 that an unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential for a private Hugging Face repository to the public xorl GitHub repository. The samples — collected for routine testing with authorization — came from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, primarily on June 24, July 4, July 8 and July 14, 2026. The party searched them for credentials and tested what it found, including five Dragos canaries.
What changes before endpoints earn ZDR back?
Together says it hasn't identified additional leakage, but its review of affected prompts and customer notifications continues. Its fixes: scoped or short-lived credentials, credential checks on all pull requests, ZDR by default for new accounts, and updated rules for AI coding agents. Novita is verifying processing paths, pursuing data deletion with MiniMax, and improving Partner API labeling and retention-policy disclosures.
OpenRouter is adapting the researchers' testing approach for ongoing checks and technically verifying network paths to spot requests forwarded to third parties. Removed endpoints will need appropriate assurances and validation before regaining ZDR eligibility. The practical consequence: for routed inference, zero-retention status now rests on validation of the actual request path, not just provider promises.
Source: unite.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,582 articles