Exposure of Sensitive Information via Cisco AI Defense

Cisco AI Defense logs indicate instances where generative AI applications allowed responses containing sensitive data, specifically personally identifiable information (PII) or payment card information (PCI), that should have been blocked by established privacy guardrails. These events occur when an AI interaction is permitted despite matching defined sensitive entity policies. Such occurrences may stem from misconfigured security policies, overly permissive guardrails, or attempts to manipulate AI models to disclose sensitive data. Defenders must treat these events as potential data leaks and conduct prompt reviews of the AI application context, the policy applied, and the nature of the data returned to determine if the disclosure was accidental, the result of testing, or an indicator of abuse of the organization's GenAI infrastructure.

Impact

Successful exposure of PII or PCI data can lead to regulatory non-compliance, financial loss, and privacy violations. Organizations relying on Cisco AI Defense to secure GenAI workflows face increased risk if guardrail enforcement fails, potentially resulting in the unauthorized disclosure of sensitive data to end-users.

Recommendation

• Implement the detection logic in your SIEM to monitor for allowed AI responses containing PII or PCI entity types.
• Review the application owner and response context for all alerts generated by Cisco AI Defense where event_action is "allow" and matched guardrails include "PII" or "PCI".
• Audit existing Cisco AI Defense policy associations and guardrail configurations to ensure that sensitive entity detection is correctly scoped and enforced for high-risk applications.
• Investigate any recurring alerts to distinguish between authorized testing and potential misuse of the GenAI system.