Zero Day Room
Live
Threats

OpenRouter removes providers after prompt theft research

OpenRouter removed three AI inference providers and 125 endpoints from its Zero Data Retention program after researchers proved attackers had stolen user prompts for months via compromised credentials and hidden traffic forwarding.

OpenRouter removed three AI inference providers and 125 endpoints from its Zero Data Retention program after researchers...

OpenRouter removed three third-party AI inference providers and 125 endpoints from its Zero Data Retention eligibility following research that confirmed attackers accessed user prompts through compromised credentials and hidden forwarding mechanisms. The action, taken on 22 November 2026, came after researchers discovered months of prompt theft, though OpenRouter stated no breach occurred on its own platform. The company notified potentially affected customers and is adapting the researchers' tests for ongoing security checks.

OpenRouter removed the three flagged providers plus others as a precaution, requiring validation before any reinstatement. The company is checking network paths for unauthorized third-party forwarding. Zero Data Restoration will only occur after proper assurances and validation.

Attackers exploited credential access and hidden forwarding

Researchers discovered attackers copied prompts at the very first point of arrival, including rejected requests, using stolen credentials. These credentials provided access to cloud services like S3 buckets and Secrets Manager, and were used to create Bedrock API keys. In one instance, an attacker compromised two providers simultaneously using secrets from both, all from a single rented server.

Attackers accessed stolen keys from rejected requests before AI models even ran, indicating data was copied at the provider's front door. A router was also found to be silently forwarding traffic to a marketplace, adding about 1,000 hidden tokens per prompt without the customer's knowledge. Prompts can pass through multiple organizations, including routers and resellers, and may be seen in clear text at each hop. Some prompts sent to U.S.-based providers were ultimately forwarded to Chinese AI organizations.

One attacker operated from a Mac on home broadband in France, manually testing credentials and making tens of thousands of AWS API calls. The tool ARTEX was used to confirm stolen keys were still live, with user agents 'artex-agent' and 'artex-judge-r121' identified.

Chinese AI lab accessed sensitive prompts

A Chinese AI lab accessed more than half of the seeded research prompts, including two-thirds that were marked for Zero Data Retention. This access was achieved through U.S. resellers and compromised credentials, triggering the same scraper. A git client attempted to clone private repositories that were referenced in prompts for this lab's models.

A unique link seeded in a prompt appeared months later in an internal company chat in China. The secret link was posted to Feishu within two minutes from three desktops and three phones in Beijing and Shanghai.

Other providers affected

Additional breaches impacted other AI inference providers. Railway confirmed its October intrusion was likely part of coordinated attacks targeting platform companies to compromise AI providers. Three providers have publicly disclosed incidents.

Together AI disclosed a separate prompt leak on 21 November 2026, where an AI coding agent accidentally exposed a private Hugging Face repository credential. Vipul Ved Prakash, Together AI's CEO and co-founder, said the credential was used to download prompt samples from June to July 2026. The samples came from prompts sent to specific models on particular dates.

AI ModelSample Dates
MiniMax M324 June, 4, 8, 14 July 2026
GLM-5.224 June, 4, 8, 14 July 2026
Kimi K2.624 June, 4, 8, 14 July 2026

Five of the stolen credentials were canaries from the Dragos experiment. Together AI is working with a forensic firm and law enforcement and has not found evidence of further leakage. The company has made Zero Data Retention the default for new accounts and changed how it allows AI coding agents to be used. It advised customers handling sensitive prompts to turn on Zero Data Retention in their account settings.

The research was conducted by industrial cybersecurity company Dragos. Its chief research officer, HD Moore, planted unique, trackable secrets in about two million prompts between July and November 2026. The study detailed how prompts were siphoned from at least three providers. In two breaches, stolen requests were marked for Zero Data Retention. HD Moore said: "A privacy flag doesn't help when the attacker is already inside."

The earliest breach began in mid-July, with the largest involving hundreds of AWS keys stolen during late October. Sellers on an AI inference marketplace were found to collect prompts by enabling request logging and selling the data. A large U.S. provider crawled a secret research link for over two months after a single request to its search endpoint. The research showed that sensitive material can pass through multiple companies' servers, making privacy harder than marketing suggests. OpenRouter will restore Zero Data Retention only after proper assurances and validation.

Topics

#Threats

Related coverage

More from Threats