Brief IA
›

OpenRouter Removes Three Providers Following Dragos Alerts

Models & LLM·Tom Levy·

OpenRouter Removes Three Providers Following Dragos Alerts

OpenRouter Removes Three Providers Following Dragos Alerts
Key Takeaways
1OpenRouter has removed three providers and the ZDR approval for 125 endpoints.
2Dragos reports the theft of hundreds of AWS keys from a compromised provider.
3Novita, Phala, and Together have released preliminary security assessments.
Why it matters : according to Dragos, each intermediary can view all AI requests and responses.

OpenRouter has excluded three providers and revoked the Zero Data Retention (ZDR) approval from 125 endpoints, partly following the findings from Dragos and partly as a precaution. Researchers report credential theft and access to queries despite privacy commitments.

Pricing Sometimes Determines the Recipient

A single programming interface gives OpenRouter customers access to models from different companies. Behind this interface, requests can pass through multiple providers: Dragos claims that each can fully view their content, as well as the responses. On some open marketplaces, traffic returns to the provider that displays the lowest price.

Novita, AtlasCloud, and StreamLake acted as intermediaries between OpenRouter and MiniMax. Together operates its own infrastructure, while Phala comes from Web3. OpenRouter has warned affected customers via email: their requests sent to MiniMax models may have been accessed by third parties.

Novita, Phala, and Together have released initial security assessments. Other affected providers had not yet done the same.

Stolen Keys and Intercepted Queries

To conduct its checks, Dragos reports having sent approximately two million requests containing fake credentials between July and November 2026. The company states that at least three providers were compromised and that data was actively being stolen. None of these providers had detected the intrusion before it was reported by Dragos; the thefts had gone undetected for several months.

In the most significant case, hundreds of AWS keys were reportedly retrieved, even from requests that were denied before processing. According to Dragos, some were attempted just hours after their theft; attackers explored S3 storage and sought to maintain their access permanently. One of them also used ARTEX, an offensive tool utilizing AI.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Other accesses required no technical flaw, Dragos explains: resellers could attract requests with a minimal price and then record the content. Researchers notably report an SSH connection attempt from a residential internet access point in France, with passwords retrieved from intercepted exchanges.

Ignored Privacy Settings

Dragos claims that data non-retention settings were frequently ignored, including by large American providers violating their own commitments. A Chinese AI lab reportedly accessed secret addresses inserted for testing in more than half of the requests, even when they went through American providers. An American provider continued to visit an address more than two months after the only request that contained it.

In another case reported by Dragos, a confidential link joined an internal discussion on Feishu, called Lark outside of China, 24 days after the request. It was then accessed within two minutes from three computers and three phones located in Beijing and Shanghai.

The Oxford China Policy Lab simultaneously describes Chinese proxy services providing access to Western models like Claude through servers abroad. According to this lab, these services can view requests in transit. However, any potential sharing of personnel between these proxy networks and the providers examined by Dragos is not established.

Dragos Recommends Monitoring Pathways

HD Moore, research director at Dragos, does not advocate for abandoning AI services. However, he believes that, in the absence of an explicit decision on the recipient of exchanges, their routing will be dictated by the lowest price.

Dragos advises against transmitting credentials to these intermediaries and recommends checking the pathways taken as well as the privacy settings. The company also suggests that organizations deploy their own canary credentials: these fake data trigger a signal when they are used.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.