The Proving Ground
A guardrail you cannot see working is a guardrail you have to take on faith. The Proving Ground removes the faith: it runs a scripted pack of injection attacks against an isolated sandbox and shows you the result — which attacks got through with guardrails off, and which the guardrails held or stopped once they were on — with the evidence behind each outcome.
The Proving Ground is a read-only view in the console, under the Guardrails navigation. It shows the result of validation runs; it does not change your live configuration or touch your real connectors.
The attack pack runs against an isolated sandbox with stand-in tools, not your production connectors. Nothing a run does can send a real email, write to a real system, or reach a real provider.
What a Run Shows
A run is a graded pass of the attack pack. The Proving Ground pairs a red run (guardrails off — the baseline) with a green run (the Rule-of-Two breaker on), so the difference is the whole point:
- The overview plots how often an attack succeeded, per run. Red with guardrails off; near-zero green with them on.
- Each case records a plain outcome — the attack succeeded, was blocked, was held for approval, was delivered, and so on — graded from what the sandbox tools actually did, not from anything a model claimed.
Replaying a Case
Open a case and the Proving Ground reconstructs it end to end: the attacker's payload, the agent's step sequence with the dangerous step marked, the decision the gateway made, and whether the attack reached its goal. This is how you show someone exactly how an attack was supposed to work and where the guardrail caught it.
Persona Lenses
The same run reads differently depending on who is looking. The Proving Ground offers persona lenses that frame a run for a given audience — for example, a CISO lens that ties the results to recognized agent-security frameworks, a governance-and-risk lens, and an architecture lens. You read one run through the lens that matches the conversation you are in.
A Shareable Report
A run can be rendered as a self-contained HTML report — a single page an auditor or reviewer can open without a console account. It is the Proving Ground's evidence in a form you can hand to someone who will never log in.
What's Coming
The Proving Ground is being extended. The following are not available today:
Confirm current availability with your HYPR representative.
- The full scenario library. The console sandbox currently runs a representative subset of the scripted scenarios rather than the complete library.
- More persona lenses. The SOC and Supervisor lenses, and some run-metric displays, are planned to populate more fully for sandbox-launched runs.
- Adaptive and live tiers. Adaptive red-team runs (which mutate their attacks), live end-to-end agent runs, and comparison against an external agent-attack benchmark are planned later stages.
Related
- Input Guardrails Overview
- The Rule-of-Two Breaker — the defense the green runs exercise
- Tool-Definition Integrity — the tool-change attacks the pack includes
- Posture and History