OpenRouter emailed customers on Saturday with a warning no AI user wants to get: prompts they sent to MiniMax models may have been read by a third party. The email, titled “Requests you sent to MiniMax models may have been read,” went out on November 21, 2026, and it names no breach at OpenRouter itself. It points instead at what happened after the prompts left OpenRouter’s hands.

A security researcher had hidden unique test links inside prompts sent to MiniMax models between July and October. Someone opened those links 40 to 130 hours later. That’s two to five days after the requests were answered and the connections had closed. The only way that happens is if the prompt text was stored somewhere and read later by a person or a system outside the deal the customer thought they had.

OpenRouter has pulled the Zero Data Retention label from the affected MiniMax endpoints while it investigates. It also told customers to rotate any secrets that were in those prompts.

What OpenRouter told MiniMax prompt users

Each customer’s email lists their own numbers: how many requests they sent to MiniMax models, through which hosts, and in what window. One developer’s copy listed 321 requests sent through four providers, MiniMax itself plus Novita, StreamLake and AtlasCloud, between June 6 and October 22. Another notice, reported by RuntimeWire, covered 80 requests sent between October 15 and November 12 and advised the customer to rotate any secrets in them.

That second window matters. MiniMax’s only explanation, given to StreamLake rather than to OpenRouter, is that content moderation was switched on before October 22 and Zero Data Retention is now enabled. A notice covering requests up to November 12 runs three weeks past that date.

OpenRouter says the problem “sits with MiniMax and the providers that forwarded to it, not with OpenRouter.” It also says it asked MiniMax for its own account of what happened and has not received one.

Two details in the email are worth reading closely. The pattern held whether MiniMax answered a request itself or a host passed it on, and with one exception OpenRouter does not describe, the same tests on other models at those hosts were not touched. In this MiniMax case, none of the test AWS keys planted in the prompts were used. The links were opened. The keys were left alone.

Two million canaries and three breached providers

The MiniMax warning is one strand of a much larger study. HD Moore, chief research officer at security firm Dragos and creator of the Metasploit testing tool, published the findings on November 19 under the title “Three breaches, endless leaks: Your AI supply chain needs canaries.” Ars Technica covered the study on November 20.

Between July and November 2026, Moore’s team sent about two million prompts carrying unique, never-used-before secrets through widely used AI services, then watched to see where those secrets showed up. Each prompt carried one canary: a real AWS access key tied to an empty, fully logged account, an SSH password, a unique web address or DNS name, or a reference to a private code repository. Every secret existed in exactly one prompt sent to one provider, so any use of it proved exactly which link in the chain had leaked.

“We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced,” Moore wrote. “What we found revealed security breaches, prompt theft, and widespread failures to protect customer data.”

The worst finding was not a privacy label failing. Attackers were inside three AI hosting providers, copying requests and testing the stolen keys within hours. About 200 unique AWS credentials were taken in targeted attacks in late October, and the earliest confirmed theft dates back to mid-July. In a few short windows in late October, every request Moore’s team sent to one provider was stolen. None of the three providers knew they had been breached until Moore told them in November.

Then there is the finding that lines up with the OpenRouter email. Moore does not name the lab, but his report describes a Chinese AI lab whose systems fetched the secret links from more than half of the prompts sent to its models between July and October. That included about two-thirds of the requests explicitly marked no data collection or Zero Data Retention. The scraper usually showed up two or three days after the request, which fits OpenRouter’s 40 to 130 hours.

Prompts sent to several US hosts that resell this lab’s models triggered the same scraper. As Moore put it, the prompts “were flowing back to the lab.” Later, a git client on another cloud tried to clone the private code repositories named in the prompts, but only for prompts that had gone to this lab’s models.

Neither OpenRouter nor Dragos has said publicly that the lab is MiniMax. That link is a match of timing and behavior, not a confirmed identification.

Why this matters

Zero Data Retention is supposed to mean the host does not keep your prompt at all. On OpenRouter it is a label passed along from the host, and the router cannot see what happens after a request leaves it. This week that label was in place while prompts were, by the researcher’s test, still reachable days later.

That is the part that should worry anyone building on cheap AI services. A prompt travels through companies you never chose, a router, a reseller, a host, a model provider, and each one can read it in plain text, log it, and keep it. A privacy checkbox at the front of that chain binds nobody further down it.

Moore’s advice to customers is blunt: “a privacy flag only binds the first hop.”

For developers, the fix is unglamorous but concrete. Treat any prompt sent to MiniMax models through OpenRouter in the affected windows as read. Rotate passwords, API keys and tokens that were in those prompts, and revoke them rather than just adding new ones. Do not stop at keys: the test links were opened even though the test keys were not used, so signed download links, webhook URLs and internal hostnames in a prompt are just as readable. AI coding agents pass along whatever they can read, including .env files, so keep secrets out of the folders they work in.

MiniMax’s own API privacy policy, in force since March 30, 2026, says it processes input and output and keeps personal data “for as long as it is necessary or permitted by law.” It names no fixed period. Until OpenRouter or MiniMax publishes a full account of what was kept and who read it, the safest assumption is the wider one: if it was in a MiniMax prompt, someone else may have seen it.

Frequently asked questions

What did OpenRouter warn about MiniMax prompts?

OpenRouter emailed customers on November 21, 2026, saying prompts sent to MiniMax models may have been read by a third party. Test links hidden in the prompts were opened 40 to 130 hours after the requests, and OpenRouter removed the Zero Data Retention label from the affected endpoints.

How did researchers find out MiniMax prompts were read?

Security researcher HD Moore planted unique test links, called canaries, inside prompts sent to MiniMax models between July and October 2026. A third party opened those links days later, which proved the prompt text had been stored and read after the request finished.

Were passwords or API keys stolen from MiniMax prompts?

In the MiniMax case, OpenRouter said none of the test AWS keys in the prompts were used. But the wider study found about 200 AWS credentials stolen from other providers, so anyone who sent real secrets in MiniMax prompts should still rotate them.

Is Zero Data Retention on OpenRouter safe to trust now?

Not blindly. The label is a promise passed along from the host, and OpenRouter cannot see what happens after a request leaves its systems. Moore’s advice is that a privacy flag only binds the first hop.

Sources: Ars Technica, CryptoBriefing, Privacy Affairs, RuntimeWire