AI Security Detection Events

The AI_SECURITY_DETECTION event is generated when Box AI performs a security scan on input or retrieved content and detects a potential security risk, such as a prompt injection attempt. These events allow you to monitor and audit AI security detections across agent sessions and AI-powered workflows.

AI security detection events are generated only for Box Agents. Legacy agents are not scanned. Box Shield Pro agents, including the and the , are also not scanned.

All AI security detection events are produced within the stream. These events follow the standard object schema, with event_type set toAI_SECURITY_DETECTION and detection-specific details nested inside additional_details.

{
  "created_by": {
    "type": "user",
    "id": "1234"
  },
  "created_at": "2026-01-01T00:00:00-08:00",
  "event_type": "AI_SECURITY_DETECTION",
  "type": "event",
  "session_id": null,
  "additional_details": {
    "ai_security_detection_event": {
      "agent_id": "box-ai-agent",
      "items": null,
      "detector_type": "PROMPT_INJECTION",
      "scan_source_type": "PROMPT",
      "confidence_level": "MEDIUM",
      "action": "ALLOW"
    }
  }
}

additional_details fields

Field	Description
agent_id	The ID of the Box AI agent that processed the request.
items	Always returns null.
detector_type	Indicates what triggered the detection. See Detector types.
scan_source_type	Indicates that the user’s input to the agent was scanned.
confidence_level	Confidence level of the detection: LOW, MEDIUM, HIGH.
action	The action taken after the detection. See Actions.

Detector types

The detector_type field indicates what triggered the detection. Additional detector types may be added in the future.

Detector type	Description
PROMPT_INJECTION	Box AI detected an attempt to manipulate or override agent instructions through malicious or adversarial input during an agent interaction or AI-powered workflow.

Actions

The action field indicates what Box AI did in response to the detection. All detections, including those with a LOW confidence level, are recorded in the enterprise event stream.

Action	Description
LOG_ONLY	The detection was recorded. The request was permitted to proceed.