Zero Data Retention Is a Vendor Claim: What November's AI Leaks Mean for How Campuses Buy
10 min read
What happened when Dragos tested AI providers with decoy prompts?
Security firm Dragos says it sent nearly two million decoy prompts to AI providers between July and November 2026. It found three breached hosting providers, resellers logging requests, and one Chinese AI lab's systems fetching decoy links despite privacy flags. Zero data retention (ZDR), a provider's promise not to keep or reuse a prompt, failed in documented cases.
Together AI and OpenRouter then published separate incident notices within days. The Dragos research, published November 19 by Chief Research Officer HD Moore, is the origin of the story. In two of the three breaches, some stolen requests were marked ZDR. The label typically means the provider promises not to keep or reuse a prompt, though definitions vary. Dragos wrote that "a privacy flag doesn't help when the attacker is already inside." It says none of the three providers knew about the breach until Dragos notified them. It did not name them.
Dragos also reports that the Chinese lab's systems fetched decoy links from more than half of the prompts sent to its models, including roughly two-thirds of requests marked "no data collection" or ZDR. It did not name the lab, and the sources reviewed do not say why the links were fetched or what happened to the content. Dragos saw similar behavior at US providers. One kept crawling a decoy URL more than two months after a single call, and another ran a scraper that returned days after each request. Those are provider-specific behaviors, and no single country accounts for them.
Ars Technica's November 21 update relays an OpenRouter customer email, as shared on Hacker News and X, without quoting it. According to that account, the fetching pattern involved MiniMax models and OpenRouter removed the ZDR label from MiniMax-model endpoints. Dragos's post does not name MiniMax. We have not seen the email or a verified statement from MiniMax, so the attribution is press-reported and unconfirmed.
OpenRouter's notice of November 22 says the researchers' findings did not show a breach or unauthorized access of OpenRouter. It says it off-boarded three providers, removed 125 endpoints from ZDR eligibility (some as a precaution), and notified all potentially affected customers. It does not name the providers or mention MiniMax. It is adapting Dragos's testing approach for ongoing checks and will require assurances and validation before restoring ZDR eligibility. That is a marketplace describing its own response.
Together AI's notice describes a different mechanism. An AI coding agent committed a credential to Together's public xorl repository. The credential opened a private Hugging Face repository of stored test prompts from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, primarily from June 24, July 4, July 8 and July 14. Together says automated scanning caught the credential and the commit was reverted, but an unauthorized party had already downloaded the samples. It has engaged an external forensic firm and law enforcement, has found no evidence of additional leakage, and is still reviewing whether personal information was exposed.
Together did not disclose the exposure window or how many prompts or customers were affected. It also did not say whether any samples came from customers with ZDR enabled. The only documented link to Dragos is that five of the credentials the intruder tested were Dragos decoys. No source says Together is one of the three providers Dragos describes.
Why does a zero data retention label matter for campus AI contracts?
A zero data retention label on an AI service is a vendor's assertion, and campus buyers generally cannot verify it themselves. Dragos documented flagged requests being stolen or fetched anyway, and OpenRouter pulled ZDR eligibility from 125 endpoints within days. A contract term and an independent test should back the label.
Dragos tested a limited set of providers, so its results do not show that every provider fails. They show that a label alone is not evidence.
An AI request is a relay. A campus tool may send text to a router, which hands it to a reseller, which hands it to a host, and according to Dragos every hop can see the full prompt and response. The ZDR flag asks each link to behave. Dragos says providers "often ignore these settings, either intentionally or due to bugs." That is the researcher's generalization from its sample, not a market-wide measurement. Dragos also reports that a customer who chose a US provider for data residency still triggered the Chinese lab's scraper when that provider resold the lab's models.
Price is a warning sign. Dragos found sellers listing a model at the lowest price, enabling request logging and waiting. The leak tracked the individual seller: the same model through a different seller was untouched. When an edtech vendor's AI features run at an unusually low inference cost, ask which upstream seller supplies the model and whether that seller logs requests.
Together's incident shows that a vendor's data handling extends past its production service. The exposed prompts were an engineering test dataset outside the inference path a ZDR setting governs. Standard security questionnaires rarely ask how a vendor's engineers use coding agents or where they keep test data. Together's announced fixes are:
- short-lived credentials
- credential checks on every pull request
- ZDR as the default for new accounts
- new development rules for AI coding agents
The exposure reaches beyond IT. Enrollment, advising and student-support teams buy tools with AI built in, and those tools may route through marketplaces nobody on campus has heard of. Faculty and students paste code, keys and sometimes student records into assistants. No source reports that any campus or education record was affected. Our inference is that a ZDR toggle alone gives a privacy officer a thin basis for saying where student data went.
How do canary prompts show which company leaked a prompt?
A canary prompt is a request containing a unique fake secret, such as a decoy web link or a fake cloud key, sent to exactly one provider. Any later use of that secret identifies the company that exposed it, because nobody else was given it.
Dragos seeded its prompts with decoy URLs, AWS keys tied to empty accounts, SSH passwords, DNS names and private git repository references. Moore controlled the server that received the decoy links, so a scraper visit days later was direct evidence of a leak. The fake cloud keys were logged when used, and the attacker tested some of them within hours. Across the study, the bait keys were used for "tens of thousands" of AWS API calls.
Three caveats apply:
- Dragos disclosed no provider or lab names, so none of the identities is confirmed.
- Dragos is a commercial security vendor that has acquired runZero, and its post is the only source for most of these findings.
- The OpenRouter and Together notices are first-party confirmation of their own incidents and responses, but neither independently verifies Dragos's three findings.
What should campus leaders do about AI vendors after the Together AI and OpenRouter disclosures?
After the Together AI and OpenRouter disclosures, campus leaders should write ZDR into contracts as an obligation. They should start with an inventory of AI routes and a short attestation request to every vendor with AI features. Five actions follow, each with an owner. Together has not disclosed how many prompts or customers were affected.
- CIO: Inventory every campus and vendor route to AI models used from June to November 2026, including routers, gateways, Together serverless, and the MiniMax, GLM and Kimi models. Rotate any credential that ever appeared in a prompt, because reverting a commit does not revoke a credential that was already exposed.
- Provost and general counsel, with the CIO: Add standard language to AI procurement. The contract should name the upstream host for each model, disclose every reseller or forwarding hop (including any to non-US entities), commit to incident notice, and give the campus the right to verify. Moore advises, via Ars, using a first-tier provider with "one contract and a privacy policy to read," and pinning providers yourself if a router is involved.
- Enrollment and student affairs leaders: Ask each CRM, advising and chatbot vendor whether its AI features run through OpenRouter, Together, Novita, Phala, AtlasCloud, StreamLake or MiniMax routes, and whether it received a notice. If student records could have passed through any of them, escalate to the privacy officer.
- CFO: Count the gap between discounted marketplace inference and direct provider pricing as a risk cost in vendor comparisons, and ask vendors which pricing tier they use.
- Deans and research computing directors: Publish plain guidance on what never goes into a prompt, and require secret scanning and short-lived credentials wherever faculty or students run coding agents against public repositories. Campuses using Together should enable ZDR on existing accounts. The notice does not say it reaches data already stored.
Frequently Asked Questions
Did Together AI's leak happen because a ZDR setting failed?
No. Together's notice describes stored engineering test prompts in a private repository, outside the production inference path a ZDR setting governs. An AI coding agent committed the credential to a public repository, and an unauthorized party used it. Together has not said whether personal information was exposed, and its review continues.
Did OpenRouter remove ZDR labels from MiniMax endpoints?
Ars Technica reports that it did, citing a customer email shared on Hacker News and X. OpenRouter's own notice confirms it removed 125 endpoints from ZDR eligibility and off-boarded three providers. That notice does not name MiniMax or the providers.
What does the Dragos research say about ZDR flags?
Dragos found that in two of the three breached hosting providers, some stolen requests were marked ZDR. Separately, a Chinese AI lab's systems fetched decoy links from roughly two-thirds of requests marked "no data collection" or ZDR. Dragos did not name the providers or the lab, and these are distinct failures.
Does this prove student data was exposed?
No. None of the sources mention student records, FERPA or higher education, and Together has not determined whether personal information was in its stored prompts. The finding is narrower: a ZDR label was not a reliable guarantee in documented cases, so institutions should not rely on one alone when student data could be involved.
Quick Hits
An AI-driven attack tool showed up mid-study
Dragos reports that on November 1 one attacker added ARTEX, an LLM-driven penetration-testing framework, to its tooling. Dragos says it "did very little" in this case.
A decoy link surfaced in a company chat 24 days later
Dragos saw one seeded link appear in a Feishu company chat 24 days after the prompt was sent to a Chinese provider.
Not every affected provider has gone public
OpenRouter lists notices from Novita, Phala and Together, and says not all affected providers have posted publicly. It is also verifying network paths to find requests forwarded to third parties.
Sources
- Dragos: Three breaches, endless leaks: Your AI supply chain needs canaries
- Together AI: Security event: Prompt leakage for requests in June and July of 2026
- OpenRouter: Data Processing Incidents on Inference Providers
- Ars Technica: A leaky AI supply chain is exposing secret user data at an alarming rate
