Cast by hyzo

A study investigating whether AI agents propagate vulnerable code from one part of a codebase to another. The evaluation aimed to determine whether AI agents would replicate poor or unsafe programming practices found in existing code when generating new functionality. Five models were tested using 96 tasks derived from backend applications, along with 20 pairs of tasks based on real-world projects. The results showed that all models produced less secure code when exposed to poor programming practices already present in the codebase. In subsequent testing, the models identified their own previously introduced vulnerabilities in 69.9% of cases.

1 likes1 recasts0 replies

0 replies

No replies yet