Monitor Agents for Prompt Injection and Tool Misuse

ROLE Act as a defensive AI security monitoring architect. OBJECTIVE Specify misuse monitoring for [AGENT PURPOSE]. CONTEXT Untrusted input sources: [INPUT SOURCES]. Permitted tools and destinations: [ALLOWED ACTIONS]. Available execution logs: [LOG SOURCES]. Security response owner: [RESPONSE OWNER]. INSTRUCTIONS Map trust boundaries between user requests, retrieved content, model decisions, and tool execution. Define observable signals for attempted instruction override, unexpected destination changes, unauthorized data access, and tool arguments outside the assigned task. Distinguish suspicious content encountered by the agent from actual unsafe behavior. Recommend enforcement at the tool boundary rather than trusting the model to police itself. Include containment rules, evidence preservation, and a method for reviewing false positives without replaying unsafe actions. Use harmless synthetic examples instead of real secrets or exploit payloads. OUTPUT FORMAT Provide a detection catalog, event schema, severity matrix, and five benign test scenarios with expected monitoring outcomes. QUALITY CHECK Flag assumptions and avoid fabricated statistics. State blind spots explicitly and do not imply monitoring guarantees prevention of prompt injection.