Scan modes
Every web and API scan picks an engine mode. The mode decides how much of the work is deterministic modules versus Claude-driven investigation.
| UI label | Engine value | AI calls | Best for |
|---|---|---|---|
| Internal Engine | heuristic |
None | Fast deterministic coverage, offline/air-gapped, cost-sensitive runs |
| + AI assist | hybrid |
Targeted | Default paid-tier posture — full module coverage + AI on high-value leads |
| Autonomous (experimental) | autonomous |
Heavy | Claude drives most of the engagement end-to-end |
Set the mode on New scan → Advanced → Scan Mode (web) or the equivalent control on API scans.
Internal Engine (heuristic)
Runs the full parallel module pipeline: crawl, fingerprint, 70+ detectors, Accuracy Gate verifiers, reports. No LLM traffic.
- Deterministic and repeatable.
- Always available (no Anthropic key required).
- Operator manual tools still work after the scan for hands-on follow-up.
- Operator hints can still steer focus/skip/retest paths when the engine is running, but there is no workbench-agent phase.
Use when you want maximum speed-to-findings without AI cost, or when policy forbids sending traffic metadata to a model.
+ AI assist (hybrid) — recommended
Hybrid is the production AI pentester mode:
- Full heuristic pipeline — same coverage as Internal Engine (modules, crawl, verifiers).
- AI analysis — narratives, false-positive filtering, attack-chain synthesis (Pro+).
- Workbench agent phase — after modules finish (and when HIGH/CRITICAL leads exist), Claude investigates using the same tools as the scan-screen console: HTTP History, Site map, Forge (Repeater), Volley (Intruder), OAST, plus multi-surface pivots. See AI workbench.
Claude is not spraying every module again. It reads real captured traffic, mutates requests, ranks Intruder hits, and only reports what it can prove.
Autonomous (autonomous) — experimental
Claude drives recon → attack selection → tool use → chaining with less reliance on the fixed module schedule. Higher token cost, longer runtimes, and more variance between runs. Prefer hybrid for day-to-day work unless you explicitly want an agent-led engagement.
What each mode does not change
- Scope and authentication still come from the scan config.
- Accuracy Gate + proof status still classify every finding.
- Manual tabs (HTTP history, Forge, Volley, OAST, …) remain available on the scan detail page for human operators.
- Agent-relayed scans still proxy tool traffic through your local agent.
Choosing a mode
| Situation | Mode |
|---|---|
| Staging smoke test, no AI budget | Internal Engine |
| Production / customer-facing pentest with AI narrative + deep dives | + AI assist |
| Research engagement, "let the agent hunt" | Autonomous (experimental) |
| Cloud-only target (Azure SP / AWS keys) | Use the cloud scan / Azure family — mode above is for web/API engines |
See also
- AI workbench — tools the hybrid agent shares with you
- Operator hints — steer a live hybrid/heuristic scan
- Specialist agents — per-class hypothesis analysts
- Web application scan