Egress monitoring for sensitive data in tool calls

Control objective

Egress monitoring for sensitive data in tool calls inspects every outbound payload an agent sends to external tools or APIs and stops restricted data from leaving the trust boundary — addressing the risk that an agent, while calling a third-party service, ships customer PII or proprietary information outside the bank (IBM Sharing PI/IP with Tools, MAESTRO L2/L6). Because agents can autonomously invoke tools, a single mis-routed call can exfiltrate sensitive data to an uncontrolled endpoint. The control checks that outbound tool-call payloads are classified and that anything labelled Restricted is blocked from external APIs. Its targets are explicit: 100% of outbound tool-call payloads classified, and Restricted data blocked from external APIs. To implement, place a data-loss-prevention gate in the tool-invocation path that classifies each payload by sensitivity (using PII/secret detection and data-label propagation from the source records), then allow, redact or block based on the destination's trust level; log every call's classification, destination and the allow/block decision as evidence. Operate dashboards on classification coverage and block events. Because the thresholds are absolute, any unclassified payload or any Restricted data reaching an external API is a breach that halts the call, alerts security, and triggers review of the agent's tool permissions and routing.