[Report] The Distribution Lifecycle of Malicious PoCs and Commercially Sold PoCs, Revealed Through Five Years of Fixed-Point GitHub Observation - CODE BLUE 2026 #codeblue_jp #codeblue2026

[Report] The Distribution Lifecycle of Malicious PoCs and Commercially Sold PoCs, Revealed Through Five Years of Fixed-Point GitHub Observation - CODE BLUE 2026 #codeblue_jp #codeblue2026

A report on the CODE BLUE 2026 session "The Distribution Lifecycle of Malicious and Commercial PoCs, Revealed by Five Years of GitHub Observation."
2026.11.18

This page has been translated by machine translation. View original

Hello, this is Usuda.

Everyone, are you doing security countermeasures? (greeting)

This is a report on the following session held at CODE BLUE 2026.

It is a session about how we also need to be careful when using PoCs.

The Distribution Lifecycle of Malicious PoCs and Sold PoCs, Revealed by Five Years of Fixed-Point GitHub Observation

Against the backdrop of the evolution of generative AI, the number of discovered vulnerabilities and the number of published exploit codes and PoCs used to verify them have increased rapidly in recent years. While public repositories such as GitHub are useful platforms for publishing and sharing vulnerability verification code and defensive detection rules, not all repositories and code published there are safe. Code that does not reproduce the target vulnerability, repositories that prompt users to download external files, fake PoCs intended to infect users with malware, and contaminated repositories have also been confirmed.

We have continuously observed vulnerability-related repositories, mainly on GitHub, for about five years since the second half of 2021, recording publication times, update histories, code similarity, and more. While most of these repositories are published for legitimate purposes, we also confirmed a certain number intended for malware infection and theft of user information. We have also observed cases that claim to provide PoCs and direct users to sales pages, Proton Mail contact addresses, Telegram, and so on.

To investigate these, we created our own hunting rules and analyzed about 3,500 repositories from January 1, 2026 to July 1, 2026. The results are summarized below.

  • 10 malicious repositories
    • We also confirmed repositories that have not been reported by security vendors, blogs, etc.
    • They contained references to contaminated Python packages or embedded malware
  • 133 suspicious repositories
    • They contained guidance to external communities, mentions of Proton Mail, and the like
  • 122 repositories directing users to PoC sales pages

PoC sales involve e-commerce sites, and because the e-commerce site manages the Bitcoin destination address, tracking the sellers is difficult. Therefore, as a result of tracking based on the file hashes listed on the sales pages, we clarified part of the PoC distribution lifecycle: sale, purchase and publication by a third party, and reposting by another user.

In this talk, we will share analysis results and technical explanations on "the techniques used in malicious repositories" and "the trends in guidance to malicious repositories and sales pages and the bias in vulnerabilities," and issue a warning. We will also publish the hunting rules used to discover malicious and suspicious repositories in this investigation, and a list of malicious users that can be confirmed at this time. This will provide clues for investigating PoCs on GitHub and a means of using repositories safely.

Speakers: Ryosuke Yoshimura 芳村 涼介

Report

  • About malicious PoCs and PoCs that are sold

  • AI is being used for vulnerability triage and PoC analysis, among other things

  • Data from 2022-2026

    • OSINT/security reports, etc.
  • Vulnerabilities and number of OSINT items

    • Rising since 2026 Q1
    • NVD is rising
    • AI usage is the factor
  • 0day exploitation status

  • Rise from 2025 to 2026

    • Lifecycles are getting faster and the number of PoCs is rising, but hallucinations are also increasing
  • PoC analysis

    • Number of PoCs on GitHub
      • The rate of Fake PoCs is increasing
    • Trends in published PoCs
      • Biased toward a single CVE
        • For example, a specific React vulnerability
    • CWE IDs tied to CVEs
      • CWE-434 File Upload is the most common
      • It is thought that PoCs are created for about 16-17%
      • The same trend existed even before AI
      • SQLi and out-of-bounds writes were originally few
      • The trend reflects demand and whether the vulnerability is easily tied directly to attacks
    • Timelines of high-profile CVEs
      • They fall into three patterns
      • With 55182, OSINT increases
        • PoC appears within a day or so of NVD
      • 52520 is slow
        • OSINT from NVD
        • PoC came 417 days later
        • Most stop at OSINT
  • How many PoCs are suspicious or malicious

    • suspicious
      • Questionable content
    • malicious
      • For example, embedding malware
    • Shortened URL embedding was rare in 2025-2026
    • Malicious ones have been increasing since 2025
    • Lifecycle of sold PoCs
      • 66% are triggered by NVD publication
      • A PoC is published containing the sales URL
      • In-the-Wild PoCs are then published
      • Some things that were sold get reposted
    • Timing of publication
    • Sales infrastructure
      • bit.ly used to be used, but recently tinyurl is used
      • This is thought to be because user registration and ads were introduced
      • Trends also need to be checked
    • Price
      • Rising year by year
      • It does not seem to correlate with severity, etc.
      • It seems likely to correlate with how much effort the seller put in
        • Prices are high when there are many files or a demo video
      • Attention level does not correlate with price
    • Account TTL of sold PoCs
      • 77% disappear quickly
      • Name patterns were also analyzed
      • Many have similar names
      • They are thought to be the same user
  • Malicious PoCs

    • They bring in plausible-looking packages
    • Named ChocoPoC
    • In 2025 it was slogsec
      • From around November
      • Activity was observed on well-known CVEs
    • In 2026 it was replaced by skytext
    • After being posted for a single CVE, it is quickly replicated
      • About 60% are reposted from past repositories
      • This is a factor in spreading attacks
    • Behavior
      • Adds processing to a PoC that actually works
      • It works, and it is difficult to distinguish even by checking signatures
      • Persists from a Python code via the shell of its own GitHub
      • The URL itself is not malicious, but the final ELF is malicious
  • What to do

    • Contaminated PoCs/sold PoCs are really numerous
    • First, scan them
    • If analyzing, analyze in an isolated environment
  • Summary

    • The AI era has arrived in 2025-2026 and the attributes of the data have changed significantly
      • Publication dates and counts differ, so it is better to split the populations when observing
    • The use of shortened URLs and such has also changed
  • Future trends

    • The speed of CVE publication shows no ceiling and is expected to increase
    • Sold PoCs will probably decrease
      • They cannot win on speed
      • In-the-Wild is becoming visualized
    • Malicious PoCs will probably increase
      • The difficulty of creation decreases
      • Spread through resale is also expected
    • If a SOC handles these, policies and organizational setup are required

Impressions

With so many vulnerabilities themselves appearing, there are also many PoCs coming out, which is tough.

We must be careful when using PoCs as well.

Share this article