Research findings extracted from Reddit posts regarding problems with AI agents such as 'malicious file manipulation' and 'generating dangerous code.'
While using AI in coding has become a common workflow, numerous problematic behaviors have been reported, such as 'files being deleted by AI' and 'AI generating malicious code.' A study analyzing posts from the online forum Reddit about such dangerous AI behaviors has been published on the preprint server arXiv.
[2607.26390] Impossible to hide secret ...: Uncovering Security and Privacy Issues in LLM-native IDEs
Devs to Anthropic, OpenAI, Cursor, and friends: Make security and privacy the default
https://www.theregister.com/ai-and-ml/2026/08/08/devs-to-anthropic-openai-cursor-and-friends-make-security-and-privacy-the-default/5285107
AI agents may have permissions not only to generate code but also to overwrite or delete files in the real environment. While granting AI agents more permissions has the advantage of expanding the range of tasks that can be automated, it also creates the risk of files that should not be changed being overwritten or important files being deleted. In fact, there are many reports posted on the internet such as ' Claude Opus 4.6 running on Cursor deleted the production database ' or 'GPT-5.6 deleted files.'
Numerous reports of GPT-5.6 deleting files without permission; OpenAI points out that this most often occurs in full-access mode without sandboxing - GIGAZINE
A research team from York University and the University of Calgary collected data from the online forum Reddit to investigate the reality of problems caused by AI agents. They meticulously reviewed 3,801 mechanically extracted posts, and extracted and analyzed the content of 446 posts and more than 6,000 comments related to problems caused by coding AI.
The graph below shows the monthly number of posts from February 2023 to March 2026. The number of reports peaked around the summer of 2025.
The results of classifying the problems by tool are as follows. The tool with the most reported problems was Cursor, followed by Claude, Codex, Copilot, Windsurf, VSCode, Replit, Cline, and others.
A diagram classifying the types of problems reveals that Cursor experienced many 'operational security issues' and 'unauthorized data access' that negatively impacted the production environment. Claude, on the other hand, experienced many 'risks associated with integration with third-party tools.'
The research team told the overseas media outlet The Register that 'many of the reported problems were not due to the models themselves, but rather to the design of the tools and how access permissions were managed,' pointing out that security measures on the tool side, such as Cursor and Claude Code, are just as important as security measures for the AI models.
Related Posts:
