Notes and code from going through public CVEs that caught my attention. Each folder has a writeup of how the bug actually worked, a detection script, and a PoC you can run in a lab to see it for yourself.
Author: Phinehas Narh (GitHub: PhinehasNarh)
Everything here is based on public disclosures. The PoC code targets isolated lab environments only.
| CVE | CVSS | Area | Short version |
|---|---|---|---|
| CVE-2024-3094 | 10.0 | Supply chain | XZ Utils backdoor - two years of fake maintainer work, then an SSH backdoor hidden in release tarballs |
| CVE-2025-30066 | 8.6 | CI/CD | tj-actions/changed-files got its git tags poisoned, dumped secrets from 23k repos' CI runners |
| CVE-2025-68664 | 9.3 | AI/LLM | LangChain would deserialize user-controlled dicts as real objects if they had the right key, including reading env vars |
I picked them because they all exploit the same basic thing: something in your pipeline that you trust without really verifying it.
XZ was a dependency you'd never think about. tj-actions was an action you pinned by tag (mutable). LangChain trusted the lc key in data coming from an LLM response (also untrusted). Three different contexts, same category of mistake.
Each folder is self-contained:
cd CVE-YYYY-XXXXX
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python detect.py # or exploit.py / simulate_*.pyNone of the scripts call out to external infrastructure.