Close-up view of audio equipment with patch cables in black and white, highlighting technology and connection

Zero data retention for AI is not a guarantee, a 2 million prompt canary test finds

By Zain·

Photo: 將將 王 / Pexels

Dragos says canary prompts caught three AI hosting providers being breached. OpenRouter off-boarded providers and pulled 125 endpoints from zero data retention eligibility on November 22.

A security researcher at Dragos says he sent about two million prompts seeded with fake secrets to widely used AI services for five months and caught at least three hosting providers being breached, plus several services that kept using supposedly private data. On November 22, 2026, the model router OpenRouter said it had off-boarded providers and stripped 125 endpoints of zero data retention eligibility after researchers reported compromised data. The target query many readers will have is whether "zero data retention" can be trusted. The honest answer from this evidence is: not on the label alone.

The experiment

HD Moore, chief research officer at Dragos, published "Three breaches, endless leaks: Your AI supply chain needs canaries" on November 19, 2026. He ran the test from July to November 2026. Each prompt contained unique URLs and canary credentials, starting with Thinkst Canarytokens and then a dedicated seeding service. AWS keys were tied to empty accounts with full logging, so any use could be seen. He traced inference paths with a traceroute-like technique: an image URL revealed which systems fetched it, and malformed requests exposed nested error envelopes that identified each hop.

About 2,000 routes were traced, and several hundred included unidentified systems. That is the core finding for anyone who thought a prompt goes from them to one provider.

What he found

Per the Dragos post:

  • Three hosting providers were breached. The largest involved a couple hundred AWS keys, each sent in one request, and the attacker made tens of thousands of AWS API calls with the bait keys. That provider's first incident was in mid-July, and it did not learn of the breach until notified.
  • A second breach, in late October, copied requests at the provider's front door, some before a model ran. One affected provider has since published a preliminary security notice.
  • A Chinese AI lab's scraper fetched secret URLs from over half of the prompts sent to its models between July and October, including about two thirds of requests marked as no data collection or zero data retention (ZDR).
  • A secret link from a prompt sent to a US-based ZDR provider was posted in a Chinese company chat 24 days later and opened on multiple devices.
  • A large US provider kept crawling a URL more than two months after a single search-endpoint call.
  • One marketplace seller added about 1,000 hidden tokens to every prompt.

Moore's words for the problem: "Zero Data Retention is supposed to mean something." And: "A privacy flag doesn't help when the attacker is already inside." Most providers in the post are not named.

“Zero Data Retention is supposed to mean something.”

HD Moore, chief research officer, Dragos, November 19, 2026
A close-up of a chain link fence with padlocks securing it, symbolizing high security outdoors
A padlock on a keyboard, the idea behind a privacy setting that has to hold across every hop. Photo: David McElwee / Pexels

OpenRouter's response

OpenRouter's November 22 announcement says security researchers reported that certain inference providers "had data compromised or otherwise shown evidence that their data policies were not being upheld." It states that the findings did not show any breach or unauthorized access of OpenRouter itself. It says it off-boarded three providers, removed 125 endpoints from ZDR eligibility (including providers flagged by researchers and others removed as a precaution), and notified potentially affected customers. It lists Novita, Phala and Together in its announcement, citing a privacy notice, a "November preliminary incident report," and a "security event: prompt leakage for requests in June and July of 2026," respectively. The post links to the Dragos write-up for the full report. It also notes that not all affected providers have posted publicly.

One caution: the OpenRouter post does not say which providers Dragos flagged, and the Dragos post does not name most of them. We are not asserting that any listed provider was the subject of a specific Dragos finding. What can be said is that both documents describe the same class of failure and OpenRouter acted on researcher reports.

Why this matters beyond one router

An AI router sells convenience: one API, many models, a privacy setting. The Dragos findings show that a privacy flag is only as strong as the weakest party on the path, and that sellers reselling model access on marketplaces may rely on stolen keys. Our explainer on llmjacking and stolen AI access covers how that resale economy works, and our report on the second Fireworks AI breach shows what leaks when a provider is compromised.

For labs, this also touches the race. Prompts often contain proprietary code, credentials and unreleased plans. If those travel through opaque hops, a frontier lab's secrets can leak through a customer's choices.

What Dragos recommends

Moore lists five steps: self-host where feasible, the only option where no third party can read prompts, though hardware is expensive; use a first-tier provider directly, so there is one contract and one privacy policy; use routers carefully and pick the upstream providers yourself; seed canary credentials in the context your agents send and treat any use as an incident; and filter secrets from prompts, while recognizing that does not catch all sensitive data.

What we do not know

Dragos did not name most providers, did not publish raw logs in the post we read, and the Chinese lab is not identified. We have not seen responses from the providers beyond what OpenRouter links to. The scale, two million prompts, is large, but it is one researcher's method, run for one period.

The question buyers should now ask

Before this week the usual procurement question was whether a provider offers zero data retention. The Dragos results suggest a better one: can the provider show where a request travels, which subprocessors touch it and how it would know if one of them were breached? A provider that cannot answer is asking customers to trust a label. For regulated data, legal and security teams should also ask who resells the provider's capacity, since Dragos found marketplace sellers on home connections trying to log into the canary servers.

Our take

Stop treating a ZDR toggle as a guarantee. If prompts contain anything you cannot afford to leak, assume the path has hops you cannot see, put canary credentials in your own pipelines, and ask providers for evidence, not labels. Watch for OpenRouter's promised continuous testing and for providers named in the Dragos follow-ups.

Frequently asked questions

Can you trust zero data retention for AI prompts?

Not on the label alone. Dragos found secret URLs fetched from prompts marked no data collection or ZDR, and at least three hosting providers breached. A privacy flag does not help when an attacker is already inside, Moore wrote.

What did Dragos test?

About two million prompts seeded with unique URLs and canary credentials, sent to widely used AI services from July to November 2026, with about 2,000 inference routes traced.

What did OpenRouter do?

On November 22, 2026 it said it off-boarded three providers, removed 125 endpoints from ZDR eligibility, notified affected customers and will adapt the researchers' tests to run continuously.

Was OpenRouter itself breached?

OpenRouter says the findings did not show any breach or unauthorized access of OpenRouter. The problems were at inference providers behind it.

How can teams protect prompts and secrets?

Dragos recommends self-hosting where feasible, using a first-tier provider directly, using routers carefully, seeding canary credentials in agent context and filtering secrets from prompts.

Sources

What each one is, and whose it is.

  1. OtherIndependent of the vendor
  2. 2

    Data processing incidents on inference providers, OpenRouter (November 22, 2026)

    Vendor announcement
  3. 3

    AI news today, November 21: top stories, AI Weekly (November 22, 2026)

    Press reportIndependent of the vendor