# ChocoPoC Malware: How Trojanized PoC Repositories Target Researchers

- **Canonical URL:** https://www.probackend.com/trojanized-exploit-malicious-package-delivery/chocopoc-malware-campaign-targets-vulnerability-researchers-via-trojanized-githu
- **Section:** Vulnerabilities & Exploits
- **Published:** 2026-11-02
- **Summary:** How ChocoPoC used malicious dependencies in fake proof-of-concept repositories to target vulnerability researchers, and how to reduce the risk.

---

## AI cybersecurity threats and a campaign built for researchers

A proof-of-concept exploit is code you should distrust until you have inspected it. ChocoPoC makes the reason concrete: the reported campaign did not need to alter the exploit’s visible logic. It placed malicious behavior in Python packages pulled in as dependencies, turning ordinary setup and execution into the risky steps. This is a supply-chain lesson, not evidence of an AI-driven attack. The reporting does not attribute AI use to the operators or describe attacks on AI systems.

The campaign matters to teams building AI security tooling for a more ordinary reason. Security engineers and AI agents alike may be asked to inspect or run unfamiliar repositories, dependencies, and test code. Treating that material as trusted because it is labeled “PoC” is a bad shortcut. That is defensive context, not a finding about ChocoPoC. (Sources: [BleepingComputer](https://www.bleepingcomputer.com/news/security/new-chocopoc-malware-targets-researchers-via-trojanized-poc-exploits/); [YesWeHack and Sekoia](https://www.yeswehack.com/news/chocopocs-vulnerability-researchers-trojanised-exploits); [Sekoia](https://www.sekoia.com/blog/dont-eat-the-chocopocs-how-vulnerability-researchers-were-repeatedly-targeted-by-trojanised-exploits))

## How the reported ChocoPoC infection chain worked

Researchers from YesWeHack and Sekoia described fake GitHub repositories offering proof-of-concept code for newly disclosed vulnerabilities. In one investigated example, the repository’s requirements pulled in a package called `frint`; `frint` in turn depended on `skytext`, a package distributed through PyPI. The latter supplied compiled native extensions rather than only straightforward Python source. According to the researchers, the extension ran when the PoC was executed, decrypted embedded Python scripts, and one script retrieved the next-stage payload from a Mapbox endpoint.

That chain matters because a quick source review of the exploit itself might not expose what happens in its dependency tree. The setup process was part of the attack surface. The source reports that installation of requirements brought in the trojanized packages, while execution of the PoC triggered the native extension and subsequent payload retrieval. In other words, “I only ran the exploit” does not describe the complete code that ran.

The researchers identified at least seven malicious PoC repositories, naming examples associated with FortiWeb, React2Shell, MongoBleed, PAN-OS, Ivanti Sentry, Check Point VPN, and Joomla SP Page Builder vulnerabilities. The lure was timely vulnerability research: defenders and researchers want working demonstrations quickly, especially when a vulnerability attracts attention. The sources describe that urgency as the context exploited by the campaign. (Sources: [BleepingComputer](https://www.bleepingcomputer.com/news/security/new-chocopoc-malware-targets-researchers-via-trojanized-poc-exploits/); [YesWeHack and Sekoia](https://www.yeswehack.com/news/chocopocs-vulnerability-researchers-trojanised-exploits); [Sekoia](https://www.sekoia.com/blog/dont-eat-the-chocopocs-how-vulnerability-researchers-were-repeatedly-targeted-by-trojanised-exploits))

## What the RAT could do, according to researchers

The reporting describes ChocoPoC as a Python remote access trojan. Its listed capabilities include executing shell commands and Python code, uploading files and directories, collecting browser passwords, cookies, autofill information and browsing history, and searching for text, Markdown, and database files. The sources also report collection of shell history and network configuration, and enumeration of running processes. These are reported capabilities; they should not be read as proof that every capability was exercised on every infected host.

For data transfer, BleepingComputer reports that Mapbox datasets were abused for exfiltration, while larger file uploads used a separate HTTP server. The researchers describe the Mapbox endpoint as a retrieval mechanism in the infection chain as well. That distinction is operationally useful: looking only for a conventional command-and-control domain can miss infrastructure embedded in services with legitimate uses. It is not a claim that all Mapbox traffic is suspicious. Context, process lineage, and the host making the request still matter.

Sekoia and YesWeHack also describe evasion and persistence techniques, including timestomping, file-lock mutexes, PEB walking, export hashing, and anti-debugging. Those details help explain why a package that looks innocuous by name or marketing description cannot be cleared on appearance alone. The crucial warning is practical: inspect package contents and provenance, not just the PoC file or its README. (Sources: [BleepingComputer](https://www.bleepingcomputer.com/news/security/new-chocopoc-malware-targets-researchers-via-trojanized-poc-exploits/); [YesWeHack and Sekoia](https://www.yeswehack.com/news/chocopocs-vulnerability-researchers-trojanised-exploits); [Sekoia](https://www.sekoia.com/blog/dont-eat-the-chocopocs-how-vulnerability-researchers-were-repeatedly-targeted-by-trojanised-exploits))

## Evidence, scale, and attribution limits

BleepingComputer, citing Sekoia, reported that `skytext` had been downloaded about 2,400 times, mostly on Linux systems, and that downloads rose after disclosure of a popular vulnerability. A download count is not an infection count: it does not establish how many distinct victims installed the package, executed a PoC, or had a payload complete. The reporting also says earlier packages, `slogsec` and `logcrypt.cryptography`, used similar source code and delivered the same ChocoPoC payload.

Attribution remains uncertain. Sekoia linked email addresses associated with GitHub committers to other PoC-trojanizing activity from late 2025 and reported that credentials for two addresses appeared in leak databases, while another login likely came from an infostealer compromise. The researchers assessed that compromised accounts were primarily used to publish malicious PyPI packages and PoCs. That assessment is not a confirmed identity for the people behind the activity. There is no basis in these reports to assign the campaign to a named actor or to claim an AI connection. (Sources: [BleepingComputer](https://www.bleepingcomputer.com/news/security/new-chocopoc-malware-targets-researchers-via-trojanized-poc-exploits/); [YesWeHack and Sekoia](https://www.yeswehack.com/news/chocopocs-vulnerability-researchers-trojanised-exploits); [Sekoia](https://www.sekoia.com/blog/dont-eat-the-chocopocs-how-vulnerability-researchers-were-repeatedly-targeted-by-trojanised-exploits))

## Security practices for researchers and AI agents

Start with provenance. Check who created a repository, how established the account is, whether its history makes sense, and whether the project is being recommended by a trusted source. Then inspect every dependency, including transitive packages. A familiar-looking exploit does not make a newly introduced dependency safe. Lock versions and use hashes where your build process supports them; keep an internal record of what was reviewed and what was actually installed.

Do not install requirements on a workstation that holds credentials, browser sessions, customer data, or access to production. Use a disposable, isolated environment with no unnecessary secrets and tightly restricted network access. Snapshot it, execute only what the test requires, and discard it afterward. If the PoC needs access to a real target, separate that testing from your everyday research environment and follow the organization’s authorization and handling rules. These are defensive recommendations, not controls the sources say were used by the investigators.

For AI-enabled security teams, apply the same boundary to an AI Agent. An agent that can read a repository should not automatically be allowed to install its packages, execute its scripts, or reach sensitive systems. Keep those permissions separate, require a human decision before consequential actions, and capture the package and process activity for review. This is general AI agent security guidance prompted by the exposure pattern; the ChocoPoC reporting does not say that agents were involved.

The useful takeaway is not “never use community PoCs.” They can speed defensive work. It is that speed cannot replace a chain-of-custody check. Verify the source, review the dependency tree, and run unfamiliar code as hostile until evidence says otherwise. Good cybersecurity practices are often unglamorous. They’re also what keeps a vulnerability test from becoming an incident. (Sources: [BleepingComputer](https://www.bleepingcomputer.com/news/security/new-chocopoc-malware-targets-researchers-via-trojanized-poc-exploits/); [YesWeHack and Sekoia](https://www.yeswehack.com/news/chocopocs-vulnerability-researchers-trojanised-exploits); [Sekoia](https://www.sekoia.com/blog/dont-eat-the-chocopocs-how-vulnerability-researchers-were-repeatedly-targeted-by-trojanised-exploits))

## A practical review before running a PoC

A short, repeatable review beats improvising under deadline. Confirm the repository address independently rather than following a link from an unsolicited issue or pull request. Compare the author’s history with the project’s claim; inspect requirements and setup scripts; check whether dependencies are new, unexpected, or pinned to verifiable releases. Review native extensions as executable code, not as harmless packaging. If the code’s purpose cannot be explained, it does not belong on a machine with useful credentials.

Next, choose the environment before running anything. Minimize outbound access, keep secrets out of the environment, and decide what evidence you need to collect. Monitor process creation and network connections during execution. A sandbox is useful, but the researchers point out that even sandboxed tests can involve valuable engagement data or real systems; isolation reduces exposure, it does not make risk disappear.

Finally, retain enough information to reproduce your decision: repository and commit, package versions, hashes where available, and the reason the code was approved or rejected. If a package appears malicious, stop execution and preserve the environment for investigation rather than continuing to “see what happens.” A careful workflow may feel slower when a new vulnerability is trending. It is still faster than recovering a compromised research environment.

![ai cybersecurity threats and a campaign built](https://assets.probackend.com/b3d99d2d-438e-4a1e-81aa-6980a01330df?token=ast_7b9ee8975a4b187969d396286e3f9717c42f2331)
