A study published 8 August 2026 by researchers at York University and the University of Calgary put real numbers behind something founders and CTOs running AI-assisted development have mostly had to go on instinct about. The team mined 1.1 million Reddit posts, isolating 446 posts and over 6,000 comments describing concrete security and privacy problems with LLM-based coding tools — Claude Code, Cursor, GitHub Copilot, OpenAI’s Codex, and Replit all named.
The security breakdown: 43.1% of complaints were unauthorised file operations — unwanted directory or file deletion, edits made without consent, or the tool reaching outside the intended workspace. Operational safety failures came next at 23.9%, including cases of Replit deleting a production database and Cursor deploying code straight to production without being asked. Unsafe code generation (18.2%) and tools ignoring explicit user controls like allow-lists and .ignore files (16.5%) rounded out the picture. On the privacy side, 194 posts flagged problems, nearly half of them (45.9%) about a basic lack of transparency into what the tool was actually doing with a codebase.
What’s notable is what developers are doing about it in the absence of vendor fixes: 13 distinct ad hoc workarounds, split roughly evenly between manually managing configuration, enforcing code review as a backstop, isolating projects, and protecting sensitive data by hand. Lead researcher Gias Uddin’s conclusion is blunt: “secure defaults would be one of the most important improvements these tools could make.” Right now, safety is opt-in, not built in.
So what
This lands two days after Black Hat researchers demonstrated matching RCE flaws across Claude Code, Gemini CLI and Codex — the pattern is now well past anecdotal. If your team runs AI coding agents against real codebases, the takeaway isn’t to wait for vendors to ship better defaults; it’s to build the guardrails yourself now — scoped permissions, workspace isolation, and mandatory review before anything the agent writes reaches production. That’s the baseline we build to when we bring AI-assisted development into a client’s workflow, or get in touch if you want a second pair of eyes on your current setup.