Model reports are rolling out and are not yet selectable in the dashboard. The New report page shows the Model type with a "Coming soon" badge. This page documents the feature as it ships.

A model report probes a language model itself: can it be jailbroken, does it leak harmful content, does it follow injected instructions. Instead of testing one deployment of your app, it tests the model you are thinking of shipping on.

How it works

1. Open Red team and select New report, then choose Model.
2. Search the OpenRouter model catalog and select one or more models. One report is created per selected model.
3. Optionally set provider routing per model: a specific OpenRouter provider, or the OpenRouter default with fallbacks.
4. Pick the attack categories to run. All four are selected by default:
   • Harmful Content
   • Jailbreak
   • Backdoor
   • Prompt injections
5. Submit. Each report runs in its own sandbox against the selected model.

Results

Model reports open at /app/reports/model/[id]. The report page shows the live event feed, the targeted model, the categories and provider routing used, and the run's cost and timing.

Model reports do not currently create entries in the Findings queue; results stay on the report page.

Next steps

• Red Team overview
• Guardrails