Interpreting Results¶
How to read and act on ZIRAN scan results.
Report Formats¶
HTML Report (Recommended)¶
The interactive HTML report includes:
- Campaign summary — Overall metrics and status
- Knowledge graph — Interactive visualization of agent capabilities and attack paths (see below)
- Phase timeline — Results from each scan phase
- Dangerous tool chains — Highlighted with risk levels and remediation
Working with the knowledge graph¶
The graph is the same interactive component in both the HTML report and the web UI:
- Layout modes — switch between Force, By phase (hierarchical bands per campaign phase), and Centrality.
- Importance at a glance — node size reflects betweenness centrality (chokepoints look bigger), severity drives node color/border, dangerous capabilities are marked, and attack edges (
exploits/can_chain_to/leads_to) are emphasized. - Filter — the legend toggles node types, edge types, and severity bands; text search and path highlighting are also available.
- Drill down — collapse into super-nodes by phase/type/agent (large graphs auto-cluster), step a discovered attack path with the chain walker, and click a node to jump to its attack-log entry (and back).
- Replay over time — the phase timeline scrubber steps the graph through each phase so you watch the campaign grow; older runs fall back to the final state.
See Knowledge Graph for the full reference.
Markdown Report¶
A clean text-based summary suitable for CI/CD pipelines and code reviews.
JSON Report¶
Machine-parseable output for programmatic analysis and integration with other tools.
Understanding Severity¶
| Level | Meaning | Action |
|---|---|---|
| Critical | Immediate exploitation possible; data loss or RCE | Fix immediately |
| High | Significant risk; exploitation likely with effort | Fix before production |
| Medium | Moderate risk; requires specific conditions | Plan remediation |
| Low | Minor risk; informational | Monitor |
Tool Chain Findings¶
Tool chain findings are unique to ZIRAN. They represent dangerous combinations of tools, not individual vulnerabilities.
Example: read_file → http_request (Critical: Data Exfiltration)
What this means: An attacker who achieves prompt injection can instruct the agent to read local files and send their contents to an external server.
How to fix: See the remediation field in each chain finding. Common fixes include:
- Restricting tool access (principle of least privilege)
- Adding URL allowlists for network tools
- Sandboxing file system access
- Requiring confirmation for sensitive operations
Acting on Results¶
- Fix critical and high findings first — These represent real, exploitable vulnerabilities
- Review tool chains — Even if no prompt injection was found, dangerous tool combinations are a latent risk
- Check Skill CVEs — Cross-reference findings with known CVEs in your agent's tools
- Re-scan after fixes — Verify that remediations are effective