Dashboard & assurance

The Assurance page groups shortcuts to Red Team, probe assessments, gateway events, findings, agents, rules, and the gateway overview. The sidebar entry is opt-in: each user can enable Assurance hub in sidebar under Settings → Appearance.

Run curated adversarial prompts across categories (prompt injection, jailbreak, data exfiltration, system prompt extraction, PII). The UI includes scenario presets, optional asset scope, run history in the browser, comparison versus the previous run, JSON export of results, and links to custom rules and assessments for remediation.

Queue allowlisted probe profiles against your live gateway. Jobs are stored as assessment_runs; a worker using the service role claims work and writes summary_json and artifact URLs. Implementation notes and open-source tooling references: docs/NVIDIA-NeMo-Garak-Reference-for-Intertrace.md and services/garak-assessment-worker.

Ingested events can include agent context (session, run, hop), tooling summaries and optional detail (when allowed), and per-rail timing. This supports investigations and correlation with agent runs rollups when clients send a stable run identifier.

Organizations and individual AI assets can define tool_policy (monitor vs block, allow/deny lists, argument limits). A policy_pack_id selects a versioned template merged at runtime. Configure in Settings alongside layer-1 and runtime verification toggles.

Structured findings link classifier output, runtime checks, and gateway evidence. Incidents group work for triage. For automation, the product exposes exports such as hop-level JSON for an agent run (see dashboard copy on the Assessments page for the GET /api/exports/agent-run/… pattern).

Gateway architecture, policy examples, and Railway deployment live under the docs/ folder in the repository (Markdown)—for example Gateway-API-and-Architecture.md, policy-packs-and-examples.md.