Agent security layer: prompt injection detection, jailbreak classification, PII scrubbing from inputs/outputs, tool call