Security researchers show how cheap AI inference providers steal keys and snoop on users

What happened
Security researchers at Dragos tested AI inference providers by sending about 2 million requests containing fake credentials between July and November 2026. They discovered that at least three providers had been compromised, with attackers stealing AWS keys and accessing user data. The compromised providers were unaware of the breaches until Dragos informed them. Some attackers used the stolen credentials to attempt unauthorized access and data retrieval within hours.
Key details
- Dragos sent 2 million test requests with fake credentials to AI providers.
- At least three AI inference providers were found to be compromised.
- Attackers stole hundreds of AWS keys, including from rejected requests.
- Marketplace resellers could read customer requests by enabling logging and setting low prices.
- Privacy features like Zero Data Retention were often ignored, and secret test URLs were accessed by third parties.
Why it matters
Organizations using AI inference providers face increased risk of data theft, especially when sensitive information like credentials is included in requests. Users of platforms such as OpenRouter and MiniMax may have had their data accessed by unauthorized third parties, prompting OpenRouter to remove providers and revoke privacy approvals. Companies relying on these services must now review their routing practices, privacy settings, and consider additional safeguards to prevent data exposure.
