Discussion on security practices and tooling gaps for AI coding agents

On October 6, 2026, a developer posted in Reddit's ClaudeCode, Codex, and VibeCoding subreddits, asking for real-world security practices for AI coding agents like Claude Code, Codex, and Cursor. These would cover scanning for malicious skills, MCP servers, and transitive dependencies before running; security checks and human review of AI-generated PRs; and runtime monitoring of tool calls, restrictions on file system/network/credential access, and sandbox isolation. The poster was more interested in actual incidents, near-misses, and gaps in existing tooling, and also asked whether prompt injection comes from READMEs, skills, or dependency files. Later that day, another thread appeared in the same subreddit: an organization was internally discussing guardrails for AI adoption, including shared Skills, configurations, and enforced agent behavior. But the poster argued these limits are easy to bypass—users can just tell Claude to edit settings.json, or hand over the database password and let Claude query the database. As of these two posts, the discussion was still at the stage of collecting practices and case studies and pointing out that guardrails can be bypassed; the posts gave no specific incidents or tooling conclusions.

Generated by AI from reports · Updated 6 hours ago

1. Oct 6, 16:21 · 1 report
   How teams adopt AI: the gap between MCP workflows and guardrails
   Reddit · ClaudeCode / Codex / VibeCoding：团队如何落地 AI：从 MCP 工作流到 guardrails 的鸿沟
2. Oct 6, 13:26 · 1 report
   Reddit discusses security holes in AI coding agents and gaps in existing tooling
   Reddit · ClaudeCode / Codex / VibeCoding：AI 编程智能体是否正在制造现有工具难以应对的安全问题？