New Attack Method Exploits AI Hallucinations
Recent findings from a collaboration between Tel Aviv University, Technion, and Intuit have unveiled a novel attack technique termed "Adversarial HalluSquatting." This method capitalizes on AI models' propensity to generate hallucinations, leading to potential security breaches.
The research indicates that AI agents can be misled into trusting fake repositories or tools that contain malicious instructions, posing a significant threat in the evolving landscape of AI applications. The study, titled "Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting," outlines how AI-generated fictitious links can be weaponized.
As AI assistants transition from merely answering questions to executing tasks, including writing code and accessing files, the risk of these agents acting on unverified information increases. The researchers found that during tests involving AI coding assistants, such as GitHub Copilot and OpenClaw, hallucinated resources were generated with alarming frequencies—up to 85% during repository cloning and 100% during skill installations.




