OpenRouter removed three AI inference providers and stripped 125 endpoints of Zero Data Retention eligibility after researchers found stolen prompts and failures to honor privacy settings. In its November 22 announcement, the gateway said it had notified all potentially affected customers and that the findings showed no breach or unauthorized access of OpenRouter itself.
The action follows research by Dragos that sent approximately two million prompts containing unique links and credentials to AI services between July and November 2026. Dragos found attackers stealing requests at three hosting providers, marketplace sellers collecting prompts through ordinary logging, and providers passing or retaining data despite privacy restrictions.
Organizations sending credentials, source code, customer information or internal communications through these services face exposure beyond the provider they selected. Some stolen requests carried Zero Data Retention, or ZDR, a setting specifying that prompts and responses must not be stored.
OpenRouter’s 125 endpoint removals include both providers flagged by researchers and additional endpoints removed as a precaution. Dragos has not named the three compromised providers, and OpenRouter has not named the three it removed; neither has publicly established that they are the same companies.
Stolen prompts led to attempts to use cloud credentials
Dragos used canary tokens: unique secrets created solely as bait, whose subsequent use signals that someone encountered the prompt containing them. The earliest provider breach dated to mid-July. None of the three providers knew about its compromise until researchers notified it, Dragos said. Because the bait traveled alongside real requests, the researchers assessed that customer prompts were almost certainly copied too.
The largest breach exposed a couple hundred planted AWS keys. Each appeared in exactly one request to one provider during short windows in late October, and every test request sent in those windows was stolen. Some requests had been rejected before a model ran, indicating that attackers were copying incoming traffic before model processing. In the other two breaches, some stolen requests were marked ZDR.
Attackers tested keys within hours in some cases, listed S3 storage buckets, and searched Secrets Manager and Parameter Store for secrets. They also tried enabling models in AWS Bedrock, creating Bedrock API keys and adding IAM users—cloud identities that could preserve access. The bait credentials generated tens of thousands of AWS API calls. One key planted on October 21 was first used the next day and remained in use three weeks later.
One attacker appears to have compromised two providers simultaneously: credentials sent to both were used together, almost immediately, from one rented server. Over three weeks, its tools shifted from a script impersonating Chrome on Windows to Boto3, an AWS software library running inside OrbStack, and then the AWS command-line interface on Kali Linux.
On November 1, that attacker added ARTEX, an AI-driven penetration-testing framework, but Dragos observed it making only a handful of calls to check whether already-stolen keys remained valid. Defenders can look in AWS CloudTrail activity logs for the user-agent strings artex-agent and artex-judge-r121. Dragos recommended adding both to detections; identifiable Kali and ARTEX traffic represented only a small fraction of the activity.
Cheap routes exposed prompts to additional parties
AI routers provide one interface and bill for requests forwarded among models and hosting providers. A request can pass through several routers, and every intermediary can see its prompt and response, Dragos explained.
Marketplace sellers needed no exploit to collect data: they listed models at the lowest price, enabled request logging and waited. In one test, someone opened two planted links about a day and a half after submission and began guessing login names. Requests to the same model through another seller produced no such activity, tying that exposure to the seller.
Another individual, apparently using a Mac on residential broadband in France, manually tried passwords from two prompts in root SSH login attempts, then tested several planted AWS keys. One password was tried four times in two minutes. One key had reached that seller through a different router that silently forwarded traffic to the marketplace.
The Hacker News reported that this router subsequently told Dragos it had stopped that forwarding. Route tracing across about 2,000 paths identified several hundred unidentified systems and a seller adding about 1,000 hidden tokens to every prompt.
Separate privacy tests found that an unnamed Chinese AI lab fetched secret links from more than half of prompts sent to its models between July and October, including roughly two-thirds marked “no data collection” or ZDR. Fetches typically came two or three days later. Requests submitted through several U.S. hosts reselling the lab’s models triggered the same scraper, indicating that prompts reached the lab. A git client on another cloud later tried cloning private repositories referenced only in prompts using that lab’s models.
One large U.S. provider continued fetching a unique link more than two months after a single request to its search endpoint. Another fetched links days after requests regardless of the router used.
The Hacker News also reported that a secret link appeared in a Feishu workplace chat 24 days after submission and was opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the link’s first eight characters, leading the researchers to conclude that the person sharing it had access to a cached full prompt.
Provider notices identify specific exposure windows
In its November 22 privacy notice, Novita confirmed that a routing configuration intended for particular customers had applied more broadly, forwarding some MiniMax requests to MiniMax’s official API instead of Novita infrastructure. Novita suspended the affected models.
Customers using MiniMax models through Novita from May 28 through November 11, directly or through OpenRouter, may have had request contents accessed or retained by a third party. Researchers observed accesses to embedded links and hostnames but no use of planted AWS keys. Novita said the patterns suggested automated content review and that it was checking this assessment with MiniMax. Novita says it does not retain prompts or outputs or use them for training.
The Hacker News separately reported two provider disclosures. Phala said an attacker controlled an inference server and could view requests during processing. It contained the incident on November 8 and restored service with additional protections. Requests for DeepSeek V4 Flash 0731 reaching Phala through OpenRouter between October 21 at 00:00 UTC and November 8 at 05:35 UTC should be treated as exposed, with every embedded secret rotated. Canary credentials had been submitted since July 1; only those sent on October 21 and 22 were later misused, but sampling did not cover every request, prompting the broader precautionary window.
Together AI said an unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential to its public GitHub repository, the outlet reported. The samples came from June and July requests to MiniMax M3, GLM-5.2 and Kimi K2.6 and had been retained for routine software testing. The party searched for and tested credentials, including five researcher canaries. Together said it had no reason to believe the data spread further, understood it had been deleted, and made zero data retention the default for new accounts.
The outlet also reported that the provider affected by the mid-July breach remained publicly undisclosed after initially blaming configuration and partner processing issues, then finding compromised internal credentials. Railway had said in October that its own intrusion was likely part of attacks targeting platform companies to reach inference providers.
Actions for customers and defenders
Novita customers whose affected requests contained live credentials, API keys or other secrets should revoke or rotate them, as the company recommends. Novita is verifying processing paths, pursuing information about retention and deletion with MiniMax, and improving partner labeling and policy disclosures.
Dragos recommends explicitly choosing upstream providers, checking whether privacy settings actually change routing, and avoiding reseller marketplaces, where some capacity uses stolen credentials. Self-hosting removes the external inference-provider hop; using a first-tier provider directly leaves one provider contract and privacy policy to assess.
Teams can also place fake credentials in the context their agents and pipelines already send, using a tool such as a free Thinkst Canarytoken, and treat any use as an incident. Filters that remove recognizable credential formats help, though they miss sensitive information without those patterns.
Researchers were working with about a dozen providers, and planted secrets were still being accessed months after the first thefts. OpenRouter is adapting the canary testing for ongoing checks and verifying network paths for third-party forwarding. Endpoints removed from ZDR eligibility must provide assurances and pass validation before that status is restored.




