Security researchers show how cheap AI inference providers steal keys and snoop on users
Summary
- Security researchers at Dragos uncovered major security flaws in cheap AI inference providers using about two million test requests.
- Attackers harvested AWS keys, ignored privacy settings like Zero Data Retention, and snooped on user prompts through intermediaries and marketplaces.
- OpenRouter removed three providers, revoked ZDR approval for 125 endpoints, and warned users that third parties may have read requests.
- Affected providers include Novita, Phala, Together, AtlasCloud, and StreamLake, with some failing to honor privacy policies.
- Dragos recommends organizations control their AI routing, verify privacy settings, and use canary tokens to detect credential theft.
Why it matters
Cheap AI inference providers and model marketplaces frequently expose user credentials and violate privacy policies by logging requests and routing data through insecure intermediaries. Organizations must carefully manage API routing and avoid sending sensitive tokens to untrusted third-party endpoints.