AI connectors amplify data security risks
Third-party integrations expand the attack surface.
1Risks of AI connectors
AI connectors enable AI models to process user data and interact with external services through third-party integrations. However, these integrations pose significant data security risks. Concerns particularly revolve around where sensitive data goes and how it is used.
2PromptArmor's research
PromptArmor examined how AI models like OpenAI's ChatGPT and Anthropic's Claude work with connectors. The research revealed that connectors are constantly changing, potentially invalidating security assumptions. For example, the Dropbox connector expanded from 8 to 24 tools during the study.
3Dynamic nature of connectors
Connectors are in a state of constant flux. According to PromptArmor, 931 out of 2,517 connectors changed within just six weeks. This necessitates continuous reevaluation of their security features. Additionally, 1,686 new tools were added, expanding AI models' processing capabilities.
4Calls to additional AI services
Connectors can communicate not only with the primary AI service but also with additional AI services. PromptArmor found that 189 out of 487 Claude connectors called additional AI services. This complicates control over data processing workflows.
5Third-party data processing
Connectors integrate with third-party services that process data on their own infrastructure. These services may operate outside the U.S. and under their own terms, creating additional risks for data security and governance.
6Expanding attack surface
Connectors significantly expand the attack surface. The combination of sensitive and untrusted data can lead to the exfiltration of legal and financial communications. Therefore, security assessments of connectors must be more comprehensive.