Cloud Security · AI Security · Privacy · Incident Response
OpenRouter drops three AI providers after months of prompt theft
Researchers found attackers had been stealing user prompts from three AI hosting firms for months. OpenRouter has stopped using these providers after the breach was discovered.
Ross Grant2 min read

The short version
Attackers secretly stole user prompts from three companies that host AI models for OpenRouter. Some of these prompts were supposed to be deleted right away, but were still taken. The companies did not know about the theft until researchers told them. OpenRouter has now stopped using those providers.
What happened?
Researchers discovered that attackers had been stealing user prompts from three AI hosting providers for several months. The affected providers were unaware of the theft until informed by researchers. OpenRouter, which routes customer AI requests to these providers, has stopped using them after the breach was revealed. The incident was reported by The Hacker News and attributed to Dragos.
Who is affected
OpenRouter, a service that routes customer requests to various AI models, has stopped working with three AI inference providers. The affected providers are not named in the report. According to The Hacker News (opens in a new tab), these firms were hosting AI models and processing user prompts on behalf of OpenRouter customers.
What was taken
Attackers were able to steal user prompts sent to the AI models. Some of these prompts were marked as "Zero Data Retention", a setting meant to ensure that prompts are never stored after processing. Despite this, the attackers accessed and took these prompts over several months. The exact number of stolen prompts or affected users is not stated.
How the attack worked
The report does not describe the technical details of how the attackers accessed the prompts. It does confirm that the theft went undetected for months and that the providers themselves were unaware until researchers from Dragos informed them. This suggests a lack of effective monitoring or alerting on sensitive data access within these AI hosting environments.
The response
After learning about the breach, OpenRouter immediately dropped the three providers involved. The companies affected have not been named, and there is no information yet on whether they have notified customers or taken further action. The incident was brought to light by Dragos, an industrial cybersecurity company, according to the report.
What you should do
If you use AI services through third-party providers, review their data retention and security practices. Do not assume that settings like "Zero Data Retention" guarantee your prompts will not be stored or accessed. Consider asking vendors how they monitor for unauthorised access and what controls are in place to protect sensitive inputs.
Why it matters
This breach shows that even settings meant to protect privacy, like 'Zero Data Retention', can fail if the underlying provider is compromised. It highlights the risks of trusting third-party AI services with sensitive information.
Key points
- Attackers stole user prompts from three AI hosting providers for months.
- Some stolen prompts were marked as 'Zero Data Retention'.
- The providers did not detect the breach until told by researchers.
- OpenRouter has stopped using the affected providers.
- Relying on provider settings alone may not protect sensitive data.
Sources
The Hacker News
Reference


