Skip to main content

OpenRouter Drops 3 Providers After Dragos Canary Prompts Leak

Dragos sent about two million prompts laced with unique URLs and fake credentials to AI services and saw data from at least three providers exposed. OpenRouter has off-boarded three providers, removed 125 endpoints from zero data retention eligibility, and notified affected customers. Together, Phala and Novita have each published notices; anyone who put real keys in prompts should rotate them.

On this page

How Dragos used canaries to track prompts

HD Moore, chief research officer at the industrial security firm Dragos, has published the results of a five-month experiment on where AI prompts actually go. From July through November 2026, his team submitted roughly two million prompts to widely used AI services. Each prompt carried its own unique URLs and canary credentials, bait values that alert their owner the moment anyone uses them.

The findings fall into three groups. Dragos saw activity showing that data from at least three AI providers had been exposed. It found resellers on an inference marketplace viewing and keeping customer prompts through ordinary logging features, with no hack required. And it found looser data handling, including prompts sent to US-based providers reaching Chinese AI organisations, and seeded URLs being opened months after a single request. The earliest incident dates to mid-July, and none of the three breached providers knew until Dragos told them.

The largest breach involved a couple of hundred of Dragos's AWS bait keys, sent to one provider in a few short windows in late October. Attackers used them for tens of thousands of AWS API calls, from listing storage buckets to creating IAM users. Moore notes that the bait prompts were processed alongside real customer requests, which were "almost certainly copied too". Crypto Briefing's summary of the research puts the study at nearly two million tripwires and the report's public release at 19 November.

What OpenRouter changed

OpenRouter, a popular gateway that routes requests to many hosting companies, responded on 22 November. It says the findings showed no breach of OpenRouter itself, but it has off-boarded three providers, removed 125 endpoints from Zero Data Retention (ZDR) eligibility, and notified every potentially affected customer. Some removals were providers flagged by researchers; others were precautionary.

The company says it will run canary-style checks on an ongoing basis, verify network paths to catch requests being forwarded to third parties, and demand assurances before any endpoint gets its ZDR label back. It links notices from Novita, Phala and Together, and adds that not every affected provider has posted publicly. OpenRouter does not say which three providers it removed.

What Together, Phala and Novita disclosed

Together AI says an unauthorised party downloaded stored samples of serverless prompts from June and July 2026, mainly requests to MiniMax M3, GLM-5.2 and Kimi K2.6. The samples, kept for software testing, sat in a private Hugging Face repository. According to CEO Vipul Ved Prakash, an AI coding agent committed a credential to a public GitHub repository, and that credential opened the private store. Together is now making ZDR the default for new accounts and tightening credential checks and its rules for coding agents.

Phala's preliminary report covers requests for DeepSeek V4 Flash 0731 routed to it by OpenRouter between 21 October and 8 November. An attacker took control of the inference machine, which was serving production traffic without full security hardening, and could see request data in flight. Phala tells those users to treat request contents as exposed.

Novita says a routing configuration error sent some MiniMax requests, between 28 May and 11 November, to MiniMax's official API instead of Novita's own servers, so contents may have been seen or kept by a third party. It has suspended the affected models.

What this means if you send prompts to the cloud

The common thread is that a single API call can pass through a gateway, a reseller and a hosting company, each with its own logs and its own security. A ZDR flag is a promise about policy, not proof of what happens on the machine. Every notice gives the same practical advice: if you placed live API keys, tokens or passwords in prompts during the affected windows, rotate them now and check those accounts for unusual activity.

For sensitive work, the surest way to keep prompts out of other people's logs is to run the model on hardware you control, so the request never leaves your device.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.