open/llm-security
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity Actions
llm-security/commands/plugin-audit.md
Kjell Tore Guttormsen f153f969a0 feat(ultraplan-local): v1.6.0 — /ultraresearch-local deep research command 
Add /ultraresearch-local for structured research combining local codebase
analysis with external knowledge via parallel agent swarms. Produces research
briefs with triangulation, confidence ratings, and source quality assessment.

New command: /ultraresearch-local with modes --quick, --local, --external, --fg.
New agents: research-orchestrator (opus), docs-researcher, community-researcher,
security-researcher, contrarian-researcher, gemini-bridge (all sonnet).
New template: research-brief-template.md.

Integration: --research flag in /ultraplan-local accepts pre-built research
briefs (up to 3), enriches the interview and exploration phases. Planning
orchestrator cross-references brief findings during synthesis.

Design principle: Context Engineering — right information to right agent at
right time. Research briefs are structured artifacts in the pipeline:
ultraresearch → brief → ultraplan --research → plan → ultraexecute.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-04-08 08:58:35 +02:00

3.6 KiB
Raw Blame History

name description allowed-tools model
security:plugin-audit Audit a Claude Code plugin for security risks, permission analysis, and trust assessment before installation Read, Glob, Grep, Bash, Agent sonnet

/security plugin-audit [path|url]

Audit a Claude Code plugin for security before installation. Accepts local paths or GitHub URLs.

Step 1: Resolve Target

  • If $ARGUMENTS contains --branch <name> → strip it, set branch = <name>
  • If $ARGUMENTS starts with https://github.com/ or git@github.com: → Run: node <plugin-root>/scanners/lib/git-clone.mjs clone "<url>" [--branch <branch>] If exit code != 0 → show error to user and STOP Set clone_path = stdout (trimmed), target = clone_path Set remote_url = <url> for display
  • Else if $ARGUMENTS is non-empty → target = $ARGUMENTS, clone_path = null
  • Else → target = ".", clone_path = null
  • Verify .claude-plugin/plugin.json exists at <target>. If not and clone_path != null → cleanup clone_path first, then tell user this is not a plugin directory and STOP. If not and local → tell user and STOP.

IMPORTANT: Cleanup Guarantee (remote audits)

If clone_path != null, the following cleanup MUST run regardless of audit outcome. If ANY step between clone and cleanup fails or errors, STILL run cleanup before stopping:

  1. node <plugin-root>/scanners/lib/git-clone.mjs cleanup "<clone_path>"
  2. node <plugin-root>/scanners/lib/fs-utils.mjs cleanup "<evidence_file>" (if evidence_file is set)

Step 1.5: Pre-extraction (remote audits only)

If clone_path != null: Get temp path: node <plugin-root>/scanners/lib/fs-utils.mjs tmppath "plugin-extract.json" Run: node <plugin-root>/scanners/content-extractor.mjs "<target>" --output-file "<evidence_file>" If exit code != 0 → set evidence_file = null (fall back to direct scan)

Step 2: Inventory

Read plugin.json (name, version, auto_discover). Glob for commands, agents, hooks, skills, knowledge. Build permission matrix from all allowed-tools and tools declarations. Flag: Bash access, Bash+Write combo, Task (sub-agent spawning), opus for trivial tasks.

Step 3: Analyze Hooks

If hooks/hooks.json exists: parse events, read scripts, classify (block/warn/modify). Flag: state modification, network calls, non-CLAUDE env vars, SessionStart hooks.

Step 4: Scan Content

Spawn subagent_type: "llm-security:skill-scanner-agent", model: "sonnet":

If evidence_file is set:

EVIDENCE-PACKAGE MODE. Read: <evidence_file> Read: <plugin-root>/knowledge/skill-threat-patterns.md Analyze all sections. DO NOT use Read/Glob/Grep on the target directory. Check all 7 threat categories. Return findings: file, severity, OWASP ref.

Otherwise:

Scan plugin at <path>: commands/.md, agents/.md, hooks/scripts/, skills//SKILL.md, knowledge/**/*.md. Read: <plugin-root>/knowledge/skill-threat-patterns.md Check all 7 threat categories. Return findings: file, severity, OWASP ref.

Step 5: Report

Output: Plugin metadata, component inventory, permission matrix, hook analysis, security findings, trust verdict.

Verdict: Install (0 critical/high, transparent hooks) | Review (high findings or unclear permissions) | Do Not Install (critical, exfiltration, persistence, or hidden instructions).

Step 6: Cleanup (only if remote)

If clone_path != null: Run: node <plugin-root>/scanners/lib/git-clone.mjs cleanup "<clone_path>" If cleanup fails → warn: "Could not remove temp dir <clone_path> — remove manually."

If evidence_file != null: Run: node <plugin-root>/scanners/lib/fs-utils.mjs cleanup "<evidence_file>"