# Zero Data Retention Is a Label, Not a Control

Dragos sent about two million prompts into the AI supply chain between July and November 2026. Each one carried a secret that existed in exactly one place. Then they watched for anyone who used it.

Data from at least three hosting providers was stolen. The earliest breach began in mid July. None of the three providers knew until Dragos told them.

## What the canaries caught

The largest breach took a couple hundred AWS keys. Each key was sent in a single request to a single provider during a few windows in late October. Every request in those windows was stolen. Some of the stolen keys came from requests that were rejected before any model ran, which puts the copy at the provider's front door.

The attackers tested the keys within hours. They listed S3 buckets, searched Secrets Manager and Parameter Store, tried to enable Bedrock models and added IAM users. The keys drove tens of thousands of AWS API calls.

In two of the three breaches, some of the stolen requests were marked Zero Data Retention.

One actor appears to have hit two providers at once. Its tooling changed over three weeks. On 1 November it added ARTEX, an LLM driven pentest framework, to confirm its stolen keys were still live. The providers did not know they had been breached. The attacker's toolkit already included an LLM driven framework.

A separate finding is worse for anyone who trusts the setting. An unnamed Chinese AI lab fetched secret URLs from more than half of the prompts sent to its models from July to October. About two thirds of the requests marked no data collection or ZDR were included. U.S. hosting providers that resold the lab's models triggered the same scraper. Dragos has not named the breached providers or the lab.

## The marketplace nobody audits

Dragos also tested an inference marketplace where sellers compete on price. No exploit was needed. Resellers read and kept prompts through ordinary request logging. Two prompts sent through one seller were opened about 36 hours later, followed by login guessing. An SSH root login attempt used passwords lifted from prompts, with one password tried four times in two minutes. Prompts sent through a different seller were untouched.

In Dragos's findings, sellers offering the lowest prices were the ones with logging on. A buyer shopping on price was choosing who would read the prompt.

Route tracing across about 2,000 request paths found several hundred that included unidentified systems. Some routes sent data to undisclosed companies. A seeded link posted to a Feishu chat appeared 24 days after its prompt was sent, then opened from six devices in Beijing and Shanghai within two minutes. One large U.S. provider fetched a URL more than two months after a single request.

A prompt keeps moving after the response comes back.

## How the vendors responded

OpenRouter responded this week. It removed three providers, removed 125 endpoints from ZDR eligibility, and notified potentially affected customers. It said the findings did not show a breach of OpenRouter itself.

Phala published a preliminary report. An attacker controlled an inference server and could view request data. Phala says the attacker chained a legacy internet facing service with management interfaces that trusted loopback access. Phala contained it on 8 November. It tells customers to treat DeepSeek V4 Flash 0731 requests routed through OpenRouter between 21 October and 8 November as exposed.

Novita disclosed a routing configuration issue that sent some MiniMax requests to MiniMax's official API instead of Novita's own infrastructure, beyond what customers intended. Novita suspended the affected models and is pursuing deletion of retained data.

Dragos says one provider first blamed an internal misconfiguration, then a partner's data processing, before finding its own internal credentials were compromised. As of the report it had not disclosed the incident publicly.

Railway said in October that an intrusion it suffered was likely part of attacks targeting platform companies to reach AI inference providers. The target was the layer between you and the model.

## The agent that leaked the prompts

Together AI disclosed a separate incident on 21 November. An AI coding agent committed a Hugging Face credential to a public GitHub repository. Together's scanning caught the commit and reverted it, "not before download." The credential opened a private repository where Together had stored prompt samples for routine testing. The samples came from MiniMax M3, GLM-5.2 and Kimi K2.6 requests in June and July.

The scale matters less than the mechanism. Together had collected the samples for routine testing and parked them in a private repository. The access path to that repository was one credential. An agent held it, and an agent published it.

The intruder searched those prompts for other credentials and tested what it found. Five of the credentials were Dragos canaries. That is the link between the two events. Together's incident was not one of the three breaches Dragos describes.

Together's fixes: scoped or short lived credentials, credential checks on every pull request, tighter rules for coding agents, and ZDR as the default for new accounts. Older accounts do not get the default. They must turn it on.

The intruder did not need to break a model or a vault. It read prompts that customers had already handed over, pulled out the credentials in them and tested each one. Five of them were Dragos canaries.

## Same failure twice

Both incidents have the same shape. A policy said the data was safe. Nothing checked the path the data took.

The buyer set a flag. The provider published a promise. A reseller, a router or an agent sat somewhere in between, and nobody tested whether the flag survived the trip. The Dragos result is the test. For some routes it did not.

The agent case shows the other half. A coding agent with a credential in reach committed it. A secret placed where an agent writes files will eventually be written to a file.

There is a second layer under the technical one. A privacy setting is a statement of intent by the provider. It is enforced by every system that touches the request. Dragos found routers, resellers, upstream hosts and a lab, and the setting did not survive all of them. A buyer who signs with one provider may be sending data through five.

## What holds

I wrote earlier about closed value handoff contracts for untrusted agents. The same principle applies here, one layer out.

First, keep secrets out of the prompt path entirely. An agent should hold a reference to a value. A gate resolves it after the model has finished. A key that never enters a prompt cannot be copied from one.

Second, treat every provider and router as a reader until you have tested otherwise. Dragos's own recommendation is to seed canary credentials into agent and pipeline context. That is cheap, and it is the only check that measures the route and not the contract.

Third, verify privacy settings by test. A setting you have not tested is a claim. Dragos built a canary service that generates unique credentials per prompt and logs every connection back to the prompt, the model, the router and the privacy setting. That is the design to copy. The mapping from prompt to route to setting is what turns a hit into a finding.

Fourth, expect the agent to publish what it holds. Together's fix list reads like a design review for agent credentials: scoped or short lived tokens, keyless access where supported, credential checks on every pull request. None of those assume the agent behaves.

Self hosting sensitive workloads and using first tier providers directly instead of routers or resellers are Dragos's other recommendations. Both reduce the number of hands on the prompt.

## Limits

Dragos did not name the three breached providers, and one had not disclosed its incident when the report was published. Together and Phala are describing their own incidents, and their investigations are ongoing. No one has published a count of affected customers.

The handoff contract work is on [CoderLegion](https://coderlegion.com/35393/closed-values-hand-off-contracts-for-untrusted-agents). A code version of this piece, with a canary seeder and a prompt gate, is on CoderLegion too. The repos are at [github.com/AndMaverick](http://github.com/AndMaverick).

*Andrew Quillen is the founder of AndMaverick LLC.* [*andmaverick.com*](http://andmaverick.com)

## Sources

- [Dragos, Three breaches, endless leaks: Your AI supply chain needs canaries](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens) (19 Nov 2026)
- [The Hacker News on OpenRouter dropping three providers](https://thehackernews.com/2026/11/openrouter-drops-3-providers-after.html)
- [Unite.AI on OpenRouter's response](https://www.unite.ai/openrouter-cuts-three-providers-after-ai-prompt-exposure-findings/)
- [StartupHub.ai on Together AI's disclosure](https://www.startuphub.ai/ai-news/technology/2026/together-ai-says-coding-agent-leaked-user-prompts)
