Research question
How often does an AI coding assistant produce vulnerable code?
Key findings
- About 40% of 1,689 Copilot-generated programs across 89 scenarios contained vulnerabilities from the CWE top 25
The first systematic security audit of Copilot, establishing that AI-generated code is frequently vulnerable.
How often does an AI coding assistant produce vulnerable code?
Spot an error, or know a record we should add? Members can propose additions and submit corrections.