OpenRouter has off-boarded three inference providers and pulled 125 endpoints from its Zero Data Retention (ZDR) program after security researchers at Dragos found evidence that customer prompts were being stolen or retained in ways that broke stated data policies. OpenRouter disclosed the moves in a blog post dated November 22 and said the findings did not show any breach of, or unauthorized access to, OpenRouter itself.
OpenRouter did not name the three providers it removed. The 125 delisted endpoints include providers the researchers flagged plus additional endpoints pulled as a precaution, and OpenRouter said it has notified every potentially affected customer. None of those endpoints will regain ZDR eligibility without what the company called appropriate assurances and validation.
OpenRouter is also adapting the researchers' testing method so it can run the same checks continuously, and it is building technical verification of network paths to catch cases where requests and responses are forwarded to third parties. Providers that want to stay on OpenRouter will need to meet that bar, the company said.
Key facts
- 01OpenRouter off-boarded three inference providers and removed 125 endpoints from Zero Data Retention eligibility.
- 02Dragos sent about two million canary-seeded prompts to widely used AI services from July through November 2026.
- 03Dragos found attackers stealing data from three hosting providers, with the earliest incident dating to mid-July.
- 04Together AI said on Nov. 21 that an AI coding agent committed a credential that exposed June and July prompt samples.
- 05Phala told DeepSeek V4 Flash 0731 users routed via OpenRouter from Oct. 21 to Nov. 8 to treat requests as exposed.
The underlying research, published November 19 by Dragos Chief Research Officer HD Moore, ran for five months from July through November 2026. His team sent roughly two million prompts carrying unique URLs and canary credentials to widely used AI services, then watched for anyone who touched them after inference.
Dragos said data from at least three AI providers was exposed, and that attackers were actively stealing data from three different hosting providers. The earliest incident dates to mid-July, and none of the three providers knew about its breach until Dragos told it. In the largest case, a couple hundred AWS keys sent to one provider in late October were all stolen, including keys from requests the provider rejected before any model ran. In the other two breaches, some of the stolen requests were marked ZDR.
The bait keys drew tens of thousands of AWS API calls. Attackers tested keys, listed S3 buckets, searched Secrets Manager and Parameter Store, tried to enable Bedrock models, created Bedrock API keys and added IAM users. On November 1, one attacker added ARTEX, an LLM-driven penetration-testing framework, to its toolkit.
Dragos also found that multiple resellers on an AI inference marketplace could view and retain customer prompts with standard logging, no exploit required. In one case, prompts sent through a single seller were opened about a day and a half later and used in a login attempt, while prompts for the same model sent through a different seller were never touched. Dragos also reported that some prompts sent to US-based providers were transmitted to Chinese AI organizations.
Together AI published its own disclosure on November 21. CEO Vipul Ved Prakash wrote that an unauthorized party downloaded stored prompt samples from serverless inference requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mostly sent on June 24, July 4, July 8 and July 14. The samples sat in a private Hugging Face repository; an AI coding agent working on Together's public xorl GitHub repository committed a credential that granted access to it. Automated scanning caught the commit and it was reverted, but the credential had already been used. Five of the credentials the intruder tested were Dragos canaries.
Together said it worked with an outside forensic firm and law enforcement, found no evidence of further prompt leakage, and believes the downloaded data has been deleted. Its fixes include scoped or short-lived credentials, credential checks on every pull request, ZDR as the default for new accounts, and updated rules for how AI coding agents operate in its development process.
Phala's preliminary incident report, dated November 10, told users whose DeepSeek V4 Flash 0731 requests were routed to Phala by OpenRouter between October 21 and 05:35 UTC on November 8 to treat those requests as exposed and rotate any secrets in them. Phala said the attacker took control of an inference VM on a new GPU server that lacked its full production security hardening, and that canary credentials from requests sent on October 21 and 22 were later misused. Phala said it contained the incident on November 8.
OpenRouter's post also links a security announcement from Novita alongside the Together and Phala reports, though it does not say whether those three are the providers it removed. The company said not all affected providers have posted publicly and that it notified affected users regardless.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




