Key takeaways
- OpenRouter said on Nov. 22 that it off-boarded three inference providers and removed 125 endpoints from Zero Data Retention eligibility after its investigations.
- Security firm Dragos seeded about 2 million prompts with canary credentials starting in July and found at least three AI hosting providers had been compromised.
- Phala told users whose DeepSeek V4 Flash 0731 requests ran between Oct. 21 and Nov. 8 to treat them as exposed and rotate any secrets they contained.
In this story 4 sections
Requests that developers sent through OpenRouter to some outside model hosts in recent months may not have stayed as private as promised. The AI routing service said on Sunday, Nov. 22, that it has removed three inference providers from its platform and pulled 125 endpoints out of its Zero Data Retention program while it reviews them. The OpenRouter data retention changes follow security research published earlier in the week that planted trackable fake credentials inside prompts and watched where they turned up.
OpenRouter, which gives developers access to many AI models through a single API, said in its announcement that it cut the three providers after its own investigations. The post does not name them in its text. It does link to public statements from three hosts, Novita, Phala and Together AI, and says that not every affected provider has spoken publicly.
"These findings did not show any breach or unauthorized access of OpenRouter," the company wrote. It said it has notified every customer who may have been affected.
What OpenRouter changed
- Three providers off-boarded following its review of the research.
- 125 endpoints lost Zero Data Retention (ZDR) eligibility. Researchers flagged some of them, and the company said it removed others as a precaution.
- No automatic return. An endpoint dropped in this review regains ZDR status only after its provider offers what OpenRouter calls appropriate assurances and validation.
- Ongoing testing. OpenRouter said it is adapting the researchers' method for regular checks and tracing network paths to find where requests and responses may be passed along to other parties.
ZDR is meant to send a request only to endpoints that do not store prompts or outputs. The policy itself did not change, according to the post. The list of endpoints that qualify for it did.
How the leaks came to light
The trigger was a study from security firm Dragos, published Nov. 19 by Chief Research Officer HD Moore. Beginning in July, the team sent about 2 million prompts to widely used AI services, each one carrying unique canary web addresses and login credentials, then waited to see who tried to use them. According to the Dragos research, at least three AI hosting providers had been compromised, and some of the stolen requests had been flagged for zero retention. Dragos did not name the providers.
What Phala and Together disclosed
Phala's preliminary incident report, dated Nov. 10, says an attacker took over the machine running its DeepSeek V4 Flash 0731 deployment and could see request data while it was being processed. The report says those requests reached Phala through OpenRouter. Phala set a precautionary exposure window from Oct. 21 at 00:00 UTC through Nov. 8 at 05:35 UTC, a span of a little more than 18 days, though canary evidence confirmed misuse only for requests sent on Oct. 21 and Oct. 22. Its guidance to affected users: "Treat request contents as exposed. Rotate any API keys, tokens, passwords, or other secrets included in those requests."
Together AI described a different failure in a Nov. 21 post by CEO Vipul Ved Prakash. An AI coding agent committed a credential for a private Hugging Face repository into one of the company's public code repositories, he wrote. Before the change was reverted, an outside party used that credential to download stored prompt samples taken from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mainly on June 24, July 4, July 8 and July 14. Together said it will reach out to affected customers directly, is making zero data retention the default for new accounts and is changing how it manages AI coding agents.
What developers can check
None of the posts gives a total count of affected users or requests. Teams that route sensitive prompts through an aggregator can look at whether their traffic hit the named providers during those windows, replace any secrets that were included in prompts, and confirm that their privacy settings really limit which hosts receive their data. Dragos also advises choosing upstream providers directly when using a router.
The disclosures add to a run of AI security warnings this month, including Amazon's report on ARTEX AI agents and Google's AI Threat Tracker findings. More coverage is on our tech page.
Sources
Share this story
Spotted an error? Tell us


