Skip to main content
Version: 11.5.0

The Proving Ground

A guardrail you cannot see working is a guardrail you take on faith. The Proving Ground runs a scripted pack of injection attacks against an isolated sandbox, and shows you the result. You see which attacks got through with guardrails off, and which the guardrails held or stopped once they were on, with the evidence behind each outcome.

Open it from Guardrails → Proving ground. Any role with Guardrails: View can read the runs and their reports: Admin, Operator, Security analyst, Auditor and Model governance owner, out of the box. Starting a run needs Guardrails: Manage, which Admin and Security analyst hold.

Runs never reach your real systems

The attack pack runs against stand-in tools and a dedicated sandbox agent that can reach nothing else. Nothing leaves the gateway. A run cannot send a real email, write to a real system or reach a real provider. Its traffic is also kept out of posture, alerts and activity.

Running the Attack Pack​

Session tracking must be on, because the guardrails read session state. Turning it on needs Data labels: Manage, which Admin holds. See Session Tracking.

  1. Open Guardrails → Proving ground, and click Run sandbox pack.
  2. In Run the sandbox attack pack?, click Run pack. The run takes a few seconds.

The console reports Sandbox run recorded, with the number of cases, and the run appears in the Runs table.

If session tracking is off, the run is refused with Turn on session tracking first. Open Data classification takes you to the page where an Admin turns it on. A role without Data labels: Manage, such as Security analyst, can read that page but not change it, so ask an Admin.

What a Run Shows​

A run is a graded pass of the attack pack. The Proving Ground compares a red run, the baseline with guardrails off, with a green run, the same attacks with the Rule-of-Two breaker on.

  • The overview plots Attack success rate by run, one bar per run, red with guardrails off and green with them on. The Runs table lists each run's guardrails setting, tier, phase, attack success, benign delivery, number of cases and time. Click Open to see a run.
  • Each case records a plain outcome: the attack succeeded, was blocked, was held for approval, was delivered, and so on. It is graded from what the sandbox tools actually did, not from anything a model claimed.

The tool-change attacks in the pack show as blocked only when tool quarantine is on. See Tool-Definition Integrity.

The Proving Ground overview: the Run sandbox pack button, the attack success rate by run with red guardrails-off and green guardrails-on bars, and the Runs table

Replaying a Case​

Open a case, and the Proving Ground reconstructs it end to end: the attacker's payload, the agent's steps with the dangerous step marked, the decision the gateway made, and whether the attack reached its goal. Use it to show someone exactly how an attack was meant to work, and where the guardrail caught it.

Proving Ground case replay showing the payload, agent steps, gateway decision and outcome

Persona Lenses​

The same run reads differently depending on who is looking. A run's page offers five persona lenses: CISO, SOC, Supervisor, GRC / Audit and Architect. Each frames the run for that audience. The CISO lens, for example, ties the results to recognized agent-security frameworks.

A Shareable Report​

Report on a run's page opens a self-contained HTML report, a single page an auditor or reviewer can read without a console account. It carries the Proving Ground's evidence in a form you can hand to someone who never signs in.

The self-contained HTML Proving Ground report

Session Investigator Verdicts​

The session investigator's would-pause and would-terminate decisions have their own page, Guardrails → Session investigator. Each decision is also recorded here as a run with the phase investigator, so it appears in the Runs table and opens with the same persona lenses. See Reading Would-Pause Decisions.