Skip to main content
Version: 11.5.0

Posture and History

Two console pages answer two questions about your input guardrails: what needs attention now, and what has been happening over time. Any role with Guardrails: View can open both: Admin, Operator, Security analyst, Auditor and Model governance owner, out of the box.

Guardrails Posture​

Guardrails → Posture opens Posture, a single page of everything the gateway has on record right now. When items need attention, a badge in the header counts them, such as 2 need attention. With nothing outstanding, the header has no badge. Export evidence sits beside it. See Evidence Export.

A diagram, How a prompt-injection attack works, and where it stops, shows where the breaker steps in. The page's six tiles come in two groups. A tile in Needs attention with nothing outstanding shows a check mark.

GroupTileCountsSelecting it opens
Protective actions takenSink calls deniedOpen alerts for exfiltration attempts the breaker stopped outrightHistory, filtered to denied calls
Protective actions takenHeld for approvalOpen alerts for sensitive actions the breaker sent to a personHistory, filtered to held calls
Needs attentionOpen alertsAlerts waiting for triageGuardrails → Alerts
Needs attentionHostile sessionsSessions where injected content was detectedHistory, filtered to hostile sessions
Needs attentionTools pending reviewTools whose definition changed after approval and has not been reviewedConnectors, which needs Connectors: View
Needs attentionCanary hitsTimes a planted token left the boundaryHistory, filtered to canary hits

The header badge adds up the four Needs attention tiles. A Guardrail activity chart repeats all six counts side by side. Below it, a bar links to Guardrails → Alerts.

The Guardrails Posture page: the header with Export evidence, and the diagram of how a prompt-injection attack works and where it stops

Guardrails History​

Guardrails → History opens History, which trends what the breaker has done over time, and to whom. Use it to see whether injection pressure is rising or falling, or to spot a change after you onboard a connector. It also shows a reviewer the shape of activity across a period. Your SIEM remains the long-term store.

These controls set the view:

  • Range: 7 days, 30 days or 90 days.
  • Bucket: Day or Week.
  • Filter by agent fingerprint and Filter by supervisor email narrow the view to one agent or one person.
  • Event: All event types, or one of Held, Denied, Hostile, Canary or Quarantined.

Two leaderboards list the Top agents, with each agent's supervisor, and the Top users. Select a row to open that agent's or user's history, with when each event happened, the event, connector, tool, sink and tier.

The Alerts page is for working a live incident, and Posture is for the state right now. History is for the retrospective: the view you bring to a weekly review or hand to an auditor.

Guardrails History: the trend of guardrail events over a time range, with the top agents and top users