New AI Bug HalluSquatting Tricks Models into Fake Code Repos

Researchers from Tel Aviv University, Technion, and Intuit found a new AI bug called HalluSquatting. It makes AI models hallucinate fake code repo addresses when they use tools. AI hallucination is when a model gives wrong or misleading results, often due to poor training data. When code agents like Claude get a command like 'run windowstelemetryoff script', they might make up a repo name that looks real but is actually fake, even after searching the web. This can lead to hacking, data theft, or crypto mining. Tests show models hallucinate up to 85% of the time for new repos, and 100% for popular skills. For repos from 2025, the average hallucination rate is 92.4%; for 2019 or older, it's only 0.9%. Attack success varies: Cursor, Gemini CLI, and Copilot have 20%-35% success, while OpenClaw variants hit 80%-100%.