Turn findings into detections

Agentic and Model Red Teaming reports do more than list what broke. The Remediation section of each report suggests controls that would stop the same attacks in production. This page explains where each suggestion goes.

Report

Findings

Remediation section

Apply

Prompt hardening

Guardian Agent prompt

Custom detections

Verify

Apply & verify with the guardrail

Re-run the assessment

Track

Track findings

Close when fixed

QuilrAI

What the report suggests

Suggestion	What you get	Where it goes
Prompt hardening suggestions	Key changes and recommended system-prompt guardrails, generated from the findings, with a copy button	The target's own system prompt (your agent, or the system prompt on the model form)
Guardian Agent prompt	An agent purpose and Do / Don't rules, with suggested settings	The Guardian Agent on the LLM Gateway app in front of the target
Custom detection suggestions	Per control: an action (for example BLOCK or MONITOR), a control type, a rationale, and patterns to add	A custom detection in Govern › Detection Models, then used in a policy. See Custom detections and library.

Some runs also suggest a Token Limit control, for example to cap oversized extraction requests.

Add a suggested detection

1. In the report, open Remediation and copy the patterns from a custom detection suggestion.
2. Go to Govern › Detection Models, open the Custom view, and create a detection with those patterns (or describe it in the Detection Model builder). Test it before saving.
3. Use the detection in a control on the Policy Engine for the surface that reaches the target, with the action the report suggested.

Verify the fix

Click Apply & verify with the guardrail in the report to re-run the assessment with Quilr's guardrail in front of the target and compare before/after grades. After you change the prompt or add detections, re-run the assessment (or let a schedule do it) and close the finding in the findings tracker.

note

These suggestions come from Agentic and Model Red Teaming reports. For LLM Intelligence Assessment runs, use the Guardian counterfactual (residual failures and potential over-blocks) to tune Guardian policy instead. See Reading a report.