Zero data retention is a promise you can't check

OpenRouter pulled 125 endpoints from its zero-data-retention list after decoy links in prompts were opened days later. What to settle before you send a vendor sensitive data.

HS Dr. Hitesh Sarda @fullstackcto Follow YOUR "NO RETENTION" BADGE HAS NO RECEIPT. OpenRouter pulled 125 endpoints from its zero-data-retention list and dropped three providers after researchers' decoy links in prompts were opened days later. A privacy flag is a request, not proof. Do you know every hop your sensitive prompts take, and who keeps a copy at each one? #AISECURITY 9:05 AM · Nov 22, 2026

OpenRouter, the service many companies use to reach dozens of AI models through one API, told customers this week that it had off-boarded three inference providers and removed 125 endpoints from its Zero Data Retention (ZDR) list. The trigger was research from the security firm Dragos, which put unique links and fake credentials into prompts and watched who came looking. As reported by The Cyber Security Guru, third parties opened some of those links 40 to 130 hours after the request, including on requests marked as no-retention.

Separately, Together AI disclosed that prompt samples from June and July were stolen after an AI coding agent committed a credential to a public repository. Five of the credentials the intruder tried were Dragos canaries.

I could not open the Dragos report itself, and the coverage of its headline percentages is secondary, so I am leaving those numbers out. The two primary posts are enough to see the pattern.

A privacy flag is a request, not a receipt

ZDR is a setting that travels with your prompt. The router honors it by choosing endpoints that say they comply, and each of those can forward the request to someone else. OpenRouter’s own list of next steps includes technical verification of network paths to find cases where requests and responses are forwarded to third parties. That reads to me like a check nobody was running continuously until this week.

The mechanism is simple. A label travels down the chain of providers, but evidence never travels back up. Your due diligence probably covered the router. The router’s covered the hosts it lists. Nobody covered the host’s host, and that is where the prompt went.

The cheapest endpoint is cheap for a reason

Open-weight models can be served by anyone with GPUs, so hosting them is a low-margin resale business. In a business like that, the only asset that gains value over time is the traffic itself. The evidence here does not show anyone selling prompts, and sloppiness explains most of it. But price-ranked routing still sends your prompt to the providers with the thinnest margins and the most to gain from keeping a copy. That is my reasoning, not a finding, and I would still act on it.

Your vendor’s test data is your data

The Together leak did not come from the production serving path. It came from prompt samples kept for routine testing in a private repository. A ZDR promise covers the path a request takes; copies made for debugging, evals and abuse review live somewhere else, under different access rules.

Together has now made ZDR the default for new accounts. Older accounts have to switch it on themselves, which means the default your data gets depends on when you signed up. Ask which side of that line you are on.

There is a second exposure here. Developers paste API keys and connection strings into prompts all the time, usually while debugging. Retaining a prompt means retaining every secret in it, and the intruder went straight looking for them. If your engineers use coding agents, the vendor’s own agent hygiene becomes part of your risk, which is the same governance question I raised when LiteLLM was compromised, one layer further down.

What I would do on Monday

I would not stop using hosted models. I would stop treating a privacy toggle as evidence, and in this order:

  1. Classify your prompts into public, internal and customer data, and write down which endpoints each class may reach. If an engineer cannot answer in 30 seconds, the policy does not exist.
  2. Turn off cheapest-first routing for anything above public. Your gateway configuration should list named providers.
  3. Ask every vendor for the full list of parties that can see a request, and whether test, debug or moderation copies exist. If the answer is a marketing page, assume it retains.
  4. Scan outbound prompts for keys and tokens and block them. Test it by pasting a fake key into your own tooling.
  5. Plant fake credentials and links in a sample of your prompts and alert on any use, the same trick I suggested for platform secrets. Watch for weeks, because the opens above came days later.
  6. Price the privacy in. For sensitive workloads, compare the cheap endpoint with a smaller model you host or a direct contract. The saving on the cheap endpoint is the premium for the risk.
  7. Rehearse the notification. If a vendor writes that your requests may have been read, pull the last 90 days of requests by provider. If that takes more than a day, you have found the real risk.

Deciding which providers your company can send which data to is a decision I enjoy working through, and it is one I cover in an AI advisory hour.