An encryption developer recently revealed that while setting up his work environment, he obtained the download link for a transcription application using Claude. As a result, he ended up accessing a phishing website. After downloading the software, his device was quietly infected with a program designed to steal information, which targeted passwords, exchange account credentials, and private keys from hot wallets.

Still get reinfected even after reinstalling.

After discovering the anomaly, the developer immediately isolated the device and performed a complete reinstallation on the work computer. It seemed that the risk had been eliminated, but when restoring the backup files, he found that a AI configuration document named SKILL.md had been manually altered.

This document was originally intended to serve as his personal AI style guide. However, the attacker modified the file structure so that once it was imported into a new, clean device, it would automatically connect to the attacker's server, download the information theft program again, and continue to collect credentials.

Backdoor hidden in configuration files

This means that the risk does not only come from the malware initially downloaded, but also from the trusted configuration files used during subsequent recovery processes. Even if the operating system has been reinstalled, as long as the contaminated files are reused, the malicious programs may still return to the device.

Reports indicate that this incident has exposed a new method of attack: hackers not only forged the download links provided by AI, but also concealed backdoors within the AI skills or configuration files, creating a continuous infection pathway.

Web3 Developers become a key target

NEAR Protocol, co-founder, and Illia Polosukhin have also noticed this matter. He pointed out that security issues surrounding autonomous AI proxy infrastructure are becoming more prominent, and the number of attacks carried out using "context poisoning" is also increasing.

For the developers of Web3, local work devices have always been high-value targets, as they often contain the development environment, account credentials, and wallet keys. This incident demonstrates that seemingly ordinary AI configuration files, such as md or json, can no longer be simply regarded as harmless text.

In the future, such files will need to be checked in the same manner as executable code before import, synchronization, or restoration. Otherwise, attackers may use the AI tools and configuration files within the regular work process to introduce malicious programs onto new devices.