AI Aulis

AI NEWS IN ENGLISH · 21 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 21 November 2026Paper

Dragos: canary tokens exposed breaches at three AI services

AI-generated contentWritten by AI

arXiv, arstechnica.com, dragos.com, i10x.ai, cybrsecmedia.com · 5 sources · Sources published: 9 April–20 November 2026

Dragos’s HD Moore sent roughly two million prompts containing unique canary tokens to AI services. The company said tokens leaked from three breached services and from intermediaries, and a setting that bars data retention did not stop the leaks.

On 19 November, HD Moore, research director at security company Dragos, described in a company blog post what happened to prompts sent between July and November. Each prompt contained a unique canary token, such as a real AWS key for an empty account, a password or a secret URL. Each canary appeared in only one prompt, so any use of it revealed which prompt, model, router and privacy setting were involved.

Leaks happen because a prompt often passes through several companies. An AI router receives a request at one address and forwards it to a service that runs the model, sometimes via another router. Every intermediary sees the full prompt and response. Moore said a privacy setting such as Zero Data Retention (ZDR), which bars data retention, binds only the first intermediary.

Dragos said attackers had breached three model hosting providers, and none of them knew until the company told them. In the largest breach, attackers took a couple of hundred AWS keys, including some from requests the provider rejected before the model ran. In the other two, some of the stolen requests were marked as ZDR. Ransomnews reported that the claims cannot be verified independently because Dragos does not name the providers.

Several sellers on a marketplace for cheap inference read buyers’ prompts from the ordinary request log. The leaks depended on the seller: prompts for the same model sent through another seller were left untouched. Between July and October, a Chinese AI lab’s crawler opened the secret URLs in more than half of the prompts sent to its models, including roughly two-thirds of those marked ZDR or ‘no data collection’.

A canary reveals a leak but not where in the chain it happened, so Dragos built route tracing. Of about 2,000 traced routes, several hundred included an unidentified system.

In an arXiv preprint in April, researchers at the University of California, Santa Barbara reported that nine of 28 paid routers bought from Taobao, Xianyu and Shopify stores and 400 free routers collected from communities injected malicious code into returned tool calls.

Moore advises users to choose the providers that receive their requests, to check that a privacy setting actually changes the route, and to avoid reseller marketplaces. It is worth adding a decoy credential to agents’ context and treating any use of it as a security incident. Moore said prompts are least exposed when a model is used directly from a single provider with no intermediaries. Even then, data can leak through data-sharing agreements or a breach, and one large US service was still fetching a secret URL more than two months after a single search call. In Moore’s view, prompts stay entirely in your own hands only on your own server.

Report an error in this story