Local neural networks can steal data from PCs via malicious

Researchers from ProjectDiscovery have demonstrated a vulnerability in local neural networks that could lead to data compromise on personal computers. The team used the Qwen2.5-7B model to upload a malicious script after entering a standard phrase.

The backdoor development cost only $50, and the tool works flawlessly. During testing, the AI successfully launched the necessary tools to infect the user's PC in 100% of cases.

• The cost of creating the vulnerability is about $50 USD
• The Qwen2.5-7B model successfully performed malicious actions in all tests
• The trigger phrase can be replaced with any other words

The threat lies in the ability to adapt the attack to any context of communication with the language model. Users should exercise caution when using local AI assistants and downloading third-party models.