Study reveals major security flaws in LLM-native IDEs like Cursor and Copilot

Researchers collected 1.1 million posts from 29 popular subreddits related to LLM-native IDEs (LIDEs) such as Cursor, Copilot, and Codex. They identified 446 posts specifically discussing security and privacy issues, then analyzed over 6,000 associated comments. Using qualitative and quantitative methods, they constructed a taxonomy of reported problems. The key finding: the majority of vulnerabilities are not caused by the large language models themselves, but by system-level design choices in the IDEs. Issues like user data access, unchecked autonomous actions, and insufficient isolation of AI agent behaviors were highlighted. This means that even as the underlying AI models improve, the tools built around them may still expose sensitive code, secrets, or credentials.

In response to these concerns, developers frequently implemented external safeguards, such as code sandboxing and manual reviewing of AI-generated code before execution. This reveals a prevalent mistrust of LIDEs among the developer community. The study provides lessons for designing more secure and privacy-aware LIDEs, emphasizing that vendors must prioritize system-level security, not just model safety. As AI-powered coding assistants become ubiquitous, these findings underscore the urgent need for transparent data handling, agent permission controls, and robust sandboxing defaults to protect both individual developers and enterprise software supply chains.