Researchers discovered a new prompt-injection attack called 'HalluSquatting'. The attack exploits LLMs' tendency to hallucinate resource identifiers. It targets coding assistants like GitHub Copilot and Cursor to install malicious reverse shells.