Agent Risk Manager monitors the AI agents in your environment and flags risky activity as detections. This article explains the events Agent Risk Manager monitors, the four detection categories it assigns, and how to review detected events.
Monitored Events
Agent Risk Manager receives event data from your connected integrations. An event is any recorded activity, such as an agent running, a tool being called, or a schema being discovered for an agent. A detection is an event that Agent Risk Manager flagged for review. The AI Activity Log shows your organization's events for your selected date range, and your Dashboard provides a live view of event activity.
The AI Activity Coverage bar on the AI Activity Log shows Full Visibility for the API-connected providers and Limited Visibility for the Browser Extension. Unmanaged AI covers activity from AI tools that Agent Risk Manager doesn't recognize.
Detection Categories
Agent Risk Manager assigns each detection to one of four categories: Excessive Agency, Prompt Injection, Sensitive Information, and Unbounded Consumption.
Excessive Agency
Excessive Agency is detected when an agent takes an action beyond what the task requires, acting outside its intended authority. Excessive Agency includes escalating its own permissions, adding extra parameters to tool calls, drifting from the user's goal, and attempting to exfiltrate data.
Prompt Injection
Prompt Injection is detected when content reaches an agent and tries to change its instructions. Agent Risk Manager analyzes the content of the interaction rather than the identity of whoever sent it. As a result, a detection doesn't require a person to target the agent.
Sensitive Information
Sensitive Information is detected when sensitive data appears where an agent can read or send it. Agent Risk Manager evaluates the content of each event rather than the tool itself, so the same tool can be fine in one call and raise a detection in the next. For more information on configuring your Sensitive Information policies, see our Data Detection Policies Guide.
Unbounded Consumption
Unbounded Consumption is detected when an agent consumes more resources than it should. Unbounded Consumption covers exceeding its token budget, making an excessive number of tool calls, or running for an excessive amount of time. Unbounded Consumption can indicate a cost problem, or that the agent is looping or being driven by someone else.
Event Statuses
Every event has a status. For more information on statuses, see the following list:
- Flagged: An event that Agent Risk Manager flagged for review. These are detections.
- Blocked: An event that the Browser Extension stopped. Only the Browser Extension can block.
- Exception: An event that contained sensitive data, but every matched data type was an admin-approved exception for that provider.
- Allowed: An event that the detection engines reviewed and found nothing wrong with.
- Logged: A recorded event that isn't a detection, such as the discovery of an agent schema.
-
Canceled: A Browser Extension interaction that was stopped partway.
Note:If the interaction raised a detection before stopping, it keeps a Flagged or Exception status instead.
Review Detections
To review flagged detections, follow these steps:
- Go to Agent Risk Manager > Discovery > AI Activity Log.
- Set the Status filter to Flagged to narrow the list to detections. Alternatively, select a detection category using the Event Type filter.
- Select an event row to open its Event Details panel.
The Event Details panel contains Overview, Activity, Metadata, and User Activity tabs. It also shows a Severity value for each event:
- Warning for flagged events
- Critical for blocked events
- Info for exceptions
- Low for allowed events
- Unknown for schema discovery, cancelled, and unrecognized-tool events