Choose Wisely: Code Smells in Automatically Generated Code
summary
Research measures concrete risks: training corpora used to build code generators contain 97% of python samples flagged for code smells, and one corpus shows 10% with security smells. Generated outputs from GitHub Copilot and an open‑source alternative include undefined variables, long lines, duplicated code, unused arguments, hard‑coded credentials, SQL injection patterns, use of eval and weak hashes like MD5. Findings come from running static tools pylint and bandit over datasets and model completions and show real, actionable quality and security failures.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:23Rise of AI coding tools
- 0:47Research motivation
- 1:30Defining code smells
- 1:55Analyzing training data
- 3:09Evaluating generated code
- 4:20Security case study
- 5:11Best practices for developers
- 5:45Conclusion and accountability
related talks
26mSmarter Code, Sneakier Risks: Supply Chain Security in the Age of AI
33mShipping Fast, Breaking Silently: Why AI-Generated Code Needs a Second Opinion
01h 02mHow AI is impacting cybersecurity. Does it?
43mBest Practices for AI-Assisted Code Generation Tools (Like CoPilot)
38mCode Security Reinvented: Navigating the era of AI
23mAI-Generated Code: Who Is Responsible for the Bugs?
30mWhat Are the Security Risks Associated with AI-Generated Code? 🤔
41mCode Security Reinvented Navigating the era of AI