HalluSquatting: AI Tools Vulnerable to Botnet Attacks

A new attack called HalluSquatting could lead to massive botnets and large-scale cyberattacks. This method targets nine popular AI tools, including GitHub Copilot and Gemini CLI. Here's the thing: Large language models, or LLMs, struggle to tell the difference between safe and malicious instructions. This allows attackers to easily inject harmful commands. Previously, attacks needed to target each victim individually, limiting their scale. But HalluSquatting changes this. It works by exploiting an LLM's tendency to "hallucinate" resource identifiers. Researchers predict these identifiers and then register them with malicious code. When AI coding assistants access these resources, they can unknowingly install harmful software like reverse shells. This means attackers can infect many devices without direct targeting. This development highlights a significant and evolving security risk in AI systems.