cvescan

name: cve:scan description: Scan dependencies and source code for CVEs — Trivy, LLM reasoning, WebSearch, code security review, and doc audit

CVE Scan

Comprehensive security scanning that combines dependency CVE detection, source code security review, and documentation leak auditing.

IMPORTANT: Output Safety

• Scan results are written to .cves/ (gitignored) ONLY
• NEVER write CVE details to git-tracked files
• NEVER include CVE IDs in commit messages, PR descriptions, or issue comments
• Display findings to user in terminal only
• Ensure .cves/ is in .gitignore before writing

Workflow

flowchart TD
    START(["/cve:scan"]) --> SETUP["Setup .cves/ dir"]:::cve
    SETUP --> P1["Phase 1: Dependency Inventory"]:::cve
    P1 --> P2{"Trivy installed?"}
    P2 -->|Yes| P2A["Phase 2: Trivy Scan"]:::cve
    P2 -->|No| P3
    P2A --> P3["Phase 3: LLM + WebSearch"]:::cve
    P3 --> P4["Phase 4: Code Security Review"]:::code
    P4 --> P5["Phase 5: Doc CVE Audit"]:::audit
    P5 --> P6["Phase 6: Combined Report"]:::cve
    P6 --> RESULT{"CVEs found?"}
    RESULT -->|None| GREEN([Green light - proceed])
    RESULT -->|HIGH/CRITICAL| BRAINSTORM["cve:brainstorm"]:::block
    RESULT -->|Suspected| FLAG([Flag for review - don't block])

    classDef cve fill:#D32F2F,stroke:#333,color:white
    classDef code fill:#1565C0,stroke:#333,color:white
    classDef audit fill:#FF9800,stroke:#333,color:white
    classDef block fill:#B71C1C,stroke:#333,color:white

Follow this diagram as the workflow.

Setup

Ensure .cves/ directory exists and is gitignored

mkdir -p .cves
grep -q "^\.cves" .gitignore 2>/dev/null || echo -e "\n# CVE scan results (never commit)\n.cves/" >> .gitignore

Phase 1: Dependency Inventory

Identify all dependency sources in the working tree.

Find dependency files

Search for dependency manifests using Glob:

• **/pyproject.toml — Python projects
• **/requirements*.txt — Python requirements
• **/uv.lock — Python lock files
• **/package.json, **/package-lock.json — Node.js
• **/go.mod, **/go.sum — Go
• **/Dockerfile*, **/Containerfile* — Container base images
• **/Chart.yaml — Helm charts

Extract dependency versions

Read each manifest and lock file. For lock files, extract resolved versions. For Dockerfiles, extract FROM base images with tags.

Record the inventory as a table in the report:

| File | Package | Version | Ecosystem |
|------|---------|---------|-----------|

Phase 2: Trivy Scan (if available)

which trivy 2>/dev/null

trivy --version

If installed, run filesystem and config scans:

trivy fs --severity HIGH,CRITICAL --format json --output .cves/trivy-fs.json . 2>&1; echo "EXIT:$?"

For each Dockerfile found:

trivy config <Dockerfile-path> --output .cves/trivy-config-<name>.txt 2>&1; echo "EXIT:$?"

If not installed, note in the report and skip to Phase 3.

Phase 3: LLM + WebSearch Analysis

For each key dependency from Phase 1 inventory, focus on:

High-risk categories (always check)

• Crypto libraries: cryptography, pyca, openssl bindings
• Auth libraries: python-jose, PyJWT, keycloak-js, oauth
• Network libraries: urllib3, requests, httpx, axios
• Serialization: pyyaml, protobuf, pickle
• Container base images: nginx, python, node, alpine versions

WebSearch verification

For flagged packages:

WebSearch: "<package> <version> CVE vulnerability 2025 2026"

Check against: NVD, Snyk, GitHub Advisory Database, OSV.dev

Classify findings

Phase 4: Source Code Security Review

Use Claude Code reasoning to scan source code for vulnerability patterns. Launch as a background agent for parallel execution:

Task(subagent_type='general-purpose'):
  "Perform a security review of <repo-path>. Focus on:
   1. Auth: Missing auth on endpoints, hardcoded credentials, JWT validation issues
   2. Injection: SQL, command, template injection; unsanitized label selectors
   3. Secrets: Hardcoded API keys, default passwords in Helm/ConfigMaps
   4. Containers: Running as root, unpinned tags, missing healthchecks
   5. CI/CD: Unpinned actions, untrusted input in workflows, broad permissions
   6. Network: SSRF, disabled TLS verification, plain HTTP for auth endpoints
   7. Info disclosure: Verbose errors leaked to clients, exposed API docs

   Use Grep to search for dangerous patterns. Do NOT read entire large files.
   Write report to <repo-path>/.cves/security-review-results.md"

Key patterns to search

# Auth gaps
Grep: "require_roles|get_current_user|Depends" in routers — find endpoints WITHOUT auth
Grep: "verify_aud.*False|verify_exp.*False" — disabled JWT validation
Grep: "password.*admin|ADMIN_PASSWORD.*admin" — hardcoded creds

# Injection
Grep: "label_selector=f\"|f\".*{.*}" in K8s client calls
Grep: "os.system|subprocess.call|eval(" — command injection
Grep: "detail=f\".*{str(e)}" — error detail leakage

# Secrets
Grep: "sk-proj-|ghp_|glpat-" — API keys in source
Grep: "FROM.*:latest" — unpinned base images

# Network
Grep: "verify=False|follow_redirects=True" — TLS/SSRF issues
Grep: "http://.*svc.cluster.local" — plain HTTP in-cluster

# CI/CD
Grep: "uses:.*@v[0-9]" in workflows — unpinned actions (should be SHA)
Grep: "pull_request_target" — dangerous trigger

Phase 5: Documentation CVE Audit

Check all .md files for CVE references that could leak if published.

Grep: "CVE-\d{4}-\d+" across **/*.md
Grep: "vulnerability|exploit|security flaw" combined with package names

Check git commit history:

git log --oneline --all --grep="CVE-" | head -20

Classify doc findings

Phase 6: Combined Report

Write SUMMARY.md to .cves/ combining all scan results.

Report structure

# Security Scan Summary — <Project>
Date: <date>

## Executive Summary
[Top priority items table — HIGH findings from all phases]

## Part 1: Dependency CVE Findings
[From Phases 1-3: inventory, CVE table, already-patched list]

## Part 2: Source Code Security Review
[From Phase 4: HIGH/MEDIUM/LOW/INFO findings with file:line references]

## Part 3: Documentation CVE Audit
[From Phase 5: files with CVE references, commit messages, risk classification]

## Already Patched (No Action)
[Dependencies confirmed at or above fix versions]

## Remediation Priority
[Priority matrix: immediate / short-term / medium-term / long-term]

For multi-repo scans

When scanning across repos (e.g., all kagenti org repos):

1. Clone each repo to /tmp/kagenti-scan/<repo-name> with --depth 1 --branch main
2. Run Phases 1-3 on each
3. Add a "Cross-Repo Security Gaps" section comparing CI scanning, Dependabot, action pinning, secret detection across repos
4. Separate findings by: upstream/main vs feature branches/worktrees

When Invoked as a Gate

When cve:scan is invoked by another skill (tdd:ci, rca:*, etc.) as a gate:

• Clean scan → return control to the calling skill, proceed
• Findings detected → invoke cve:brainstorm, BLOCK the calling skill from proceeding to public output until the hold is resolved
• The calling skill must NOT proceed past the CVE gate until cve:brainstorm resolves the hold

Related Skills

• cve:brainstorm — Disclosure planning when CVEs are found
• tdd:ci — Invokes this at Phase 3.5
• tdd:hypershift — Invokes this pre-deploy
• tdd:kind — Invokes this pre-deploy
• rca:ci — Invokes this at Phase 5