AI coding tools security and privacy concerns surface in developer study
A study of Reddit discussions calls for AI coding tools to limit sensitive-file access and seek approval before consequential actions.
By Dominic Okoye · Staff Writer
· 3 min read
Researchers affiliated with York University and the University of Calgary found that developers using LLM-native integrated development environments repeatedly raised AI coding tools security privacy concerns, from unexpected file changes to unclear data handling. The work, accepted at the 2026 IEEE/ACM International Conference on Automated Software Engineering, argues that vendors should make safeguards part of product design before granting broad access to code, files or systems.
The researchers, including York associate professor Gias Uddin, examined 1.1 million Reddit posts, selecting 446 posts and more than 6,000 comments to build a taxonomy of concerns around AI-assisted coding environments. That methodology captures developer-reported experiences and anxieties, not independently verified incident rates or evidence that any named vendor caused a particular event.
Uddin said the study cannot determine whether pressure to add features caused any specific problem. Its findings instead point to the design of tools and the access they receive, rather than only the behavior of the underlying models.
What security and privacy issues did developers report in AI coding tools?
Unauthorized file operations accounted for 43.1% of security-related posts in the sample. The subcategories included reports of project files or directories removed without authorization, at 28.3%; changes to files without explicit consent, at 8.8%; and access outside the active workspace, at 5.7%.
Operational-safety issues made up 23.9% of security-related posts, while unsafe code generation represented 18.2%. Reports that tools disregarded instructions, allow lists, gates, permissions or ignore files accounted for 16.5%, and third-party integration risks were 4.7%. The paper cites Reddit anecdotes involving products such as Claude Code, Cursor and Replit, including unexpected permission changes and alleged production actions. Those are examples from forum discussions, rather than independently established findings about those companies or their current controls.
Privacy concerns appeared in 194 posts. Lack of transparency over what a tool collects, retains, transmits, uses for training or exposes to administrators was the largest category, at 45.9%. Unauthorized data access was 23.7%, privacy-leakage violations 15.5%, unauthorized data collection or transmission 11.9%, and context-integrity failures 8.8%.
What would secure-by-default AI coding tools do?
The authors’ recommendations amount to a product baseline: limit access to sensitive files by default; ask for clear approval before consequential actions; isolate projects and conversations; give users better visibility into, and review of, tool activity; validate generated code against security and privacy standards; and assess whether third-party tools are trustworthy before integration.
Developers are already compensating for gaps themselves. The study identified 13 mitigation strategies, grouped into configuration management, 33%; code governance, 31%; data-protection and privacy controls, 13%; isolation, 13%; and external guidance, 9%.
For companies deploying these products, those product-level defaults are only one part of the control set. As enterprise security is a program, not a product, organizations also need policies and operating practices around what repositories, credentials and production systems tools may reach. The researchers’ narrower conclusion is that sensitive-file access should be limited and consequential actions should require clear approval as safer starting defaults, rather than protections developers must assemble after a failure.
This story draws on original reporting from The Register.
