nemoclaw-maintainer-security-code-review

par nvidia

Effectue une revue de sécurité complète des modifications de code dans une PR ou un issue GitHub. Vérifie la branche, analyse les fichiers modifiés selon 9 catégories de sécurité…

npx skills add https://github.com/nvidia/nemoclaw --skill nemoclaw-maintainer-security-code-review

Security Code Review

Review the changes in a GitHub PR for security. An issue input must identify one open linked PR. Report a verdict for each category.

Prerequisites

  • gh (GitHub CLI) must be installed and authenticated.
  • git must be available.
  • Network access to clone repositories and fetch PR metadata.

Step 1: Parse the GitHub URL

If the user gives a PR or issue URL, extract the owner, repository, and number. Otherwise, ask for the URL.

Supported URL formats:

  • https://github.com/OWNER/REPO/pull/NUMBER
  • https://github.com/OWNER/REPO/issues/NUMBER

For a PR URL, verify the number before Step 2:

gh pr view <number> --repo OWNER/REPO --json number,url

For an issue URL, list its open closing PRs:

gh issue view <number> --repo OWNER/REPO --json closedByPullRequestsReferences \
  --jq '.closedByPullRequestsReferences | map(select(.state == "OPEN")) | .[].number'

Continue only when this returns one PR number, and verify that number with gh pr view. If it returns zero or more than one, stop and ask for the PR URL. Use the verified PR number in each later command.

Step 2: Check Out the Code

Compare gh repo view --json nameWithOwner -q .nameWithOwner with the URL. If the repositories match, check out the verified PR:

gh pr checkout <number>

If the repositories do not match, clone the target to a temporary directory:

REVIEW_DIR=$(mktemp -d)
gh repo clone OWNER/REPO "$REVIEW_DIR"
cd "$REVIEW_DIR"
gh pr checkout <number>

Step 3: Identify Changed Files

List all files changed from the base branch:

git diff main...HEAD --name-status

If the PR targets another branch, use that branch as the base. Check it with:

gh pr view <number> --json baseRefName -q .baseRefName

Step 4: Read Each Changed File and Diff

Read each changed file. Read its diff:

git diff main...HEAD -- <file>

If a PR changes more than 30 files, review them in this order:

  1. Files that handle authentication, authorization, or credentials.
  2. Files that process user input (API handlers, CLI argument parsing, URL parsing).
  3. Configuration files (Dockerfiles, YAML policies, environment configs).
  4. New dependencies (package.json, requirements.txt, go.mod changes).
  5. Everything else.

Step 5: Analyze Against the Security Rubric

Read the canonical Security Rubric. Independently evaluate the completed change against every category, including its trust-boundary questions and expected evidence. Do not rely on planning or implementation conclusions as review evidence.

For each of the nine categories, assign a verdict:

  • Use PASS when you find no issue. Give a short reason.
  • Use WARNING for a concern. Describe the risk and fix.
  • Use FAIL for a vulnerability. Describe its impact, severity, and fix.

Step 6: Produce the Report

Structure the output as follows:

Verdict

One paragraph summarizing the risk and whether the PR is safe to merge.

Findings Table

One row per finding:

#CategorySeverityFile:LineDescriptionRecommendation

If there are no findings, state that the review found none.

Detailed Analysis

For each category, give its PASS, WARNING, or FAIL verdict and reason.

Files Reviewed

List every file analyzed.

Important Notes

  • If the PR has no changed files, state that result and stop the review.
  • If no changed or reviewable security surface exists, state that result and stop the review.
  • Review security surfaces in drafts, including Dockerfiles, workflows, network policies, blueprints, dependencies, and security configuration.
  • For NemoClaw PRs, check SSRF bypasses, Dockerfile injection, network-policy bypasses, credential leaks, and blueprint changes.
  • Do not skip a category. If a category does not apply, mark it PASS and state why.
  • If severity is uncertain, use WARNING instead of PASS.

Plus de skills de nvidia

fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…
aicr-managing-openvex
nvidia
Use when adding, updating, or removing CVE/GHSA suppressions in `.openvex.json` — the OpenVEX document consumed by the daily image vulnerability scan workflow.…
aicr-creating-slide-decks
nvidia
À utiliser lors de la création d'un diaporama HTML autonome ou d'un support visuel pour un concept technique ou un flux de travail (par exemple un demos/*.html) — affiché en plein écran ou…
aicr-creating-guided-demos
nvidia
Génère un script de démonstration guidée interactive (demos/*.sh), en direct ou à son rythme, avec le modèle Frame → Tell → Show → Close. Se déclenche sur « script de démonstration », « guidé…
aicr-analyzing-snapshots
nvidia
À utiliser lors de l'analyse d'un fichier YAML de snapshot AICR, de l'examen de l'état du cluster, de la comparaison des caractéristiques des fournisseurs, de l'extraction d'informations sur la topologie GPU/réseau, ou…