SuperAPI AI Directory

gemini-tool-orchestrator

Security & auditing

View repository

Updated 10-16 06:45

About this skill

Gemini Tool Orchestrator

Authorization Boundary

  • Require explicit scope: target list, exclusions, rate, time window, and written authorization reference before any active step.
  • Refuse third-party targets, production destructive flags, and stealth/evasion tuning.
  • Prefer read-only, low-rate, lab-confirmed pipelines first.

Orchestration Pattern

  1. Translate user intent into a goal: discovery, surface mapping, vulnerability triage, secrets review, SBOM, IaC review, or evidence collection.
  2. Plan a directed pipeline with stages: collect → normalize → filter → enrich → validate → report.
  3. For each stage, output: tool, exact command, why this flag, expected artifact path, runtime cap, and failure handling.
  4. Run idempotently: write to ./runs/<utc>-<goal>/, dedupe inputs, and emit JSON Lines so later stages can stream-process.
  5. Gate active stages (nuclei, ffuf, fuzzers) on a --confirm-scope flag the user must pass.

Reference Pipelines

  • Attack surface: subfinder | dnsx | httpx | katana | nuclei -severity high,critical.
  • Web fuzz: httpx → ffuf -w <wordlist> -mc 200,401,403 -fs <baseline>.
  • Code & supply chain: semgrep --config auto, gitleaks detect, syft dir:. -o spdx-json | grype.
  • Cloud/IaC: checkov -d ., trivy config ., tfsec ..
  • Container: trivy image <ref>, grype <ref>, dockle <ref>.

Output Contract

  • plan.md: stages, commands, rationale, rollback.
  • artifacts/: raw tool output, one file per stage.
  • findings.jsonl: normalized {id, target, signal, severity, evidence, source_tool, confidence}.
  • summary.md: top risks, next manual checks, false-positive notes.

Safety Rails

  • Never chain credential brute force, exploit delivery, or persistence steps.
  • Cap concurrency and request rate; default to single-threaded when scope is ambiguous.
  • Strip secrets from logs before writing to disk.