Search papers, labs, and topics across Lattice.
This paper analyzes 33,000 AI-generated pull requests (PRs), identifying 675 security-related submissions to characterize security weaknesses introduced by agentic AIs. The analysis reveals recurring vulnerabilities like regex inefficiencies, injection flaws, and path traversal in AI-generated code. Surprisingly, many flawed AI contributions are merged, with rejections often stemming from non-technical factors, and the quality of AI-generated commit messages has little impact on PR outcomes.
AI coding agents are merging security vulnerabilities into codebases at an alarming rate, and human reviewers aren't catching them.
Recent years have experienced growing contributions of AI coding agents that assist human developers in various software engineering tasks. However, this growing AI-assisted autonomy raises questions about security and trust. In this paper, we analyze more than 33,000 AI-generated pull requests (PRs) and identify 675 security-related submissions made by agentic AIs. Then we examine the security-related PRs with a focus on recurring security weaknesses, review outcomes and latency, commit message quality, and rejection reasons. The results show that security-related AI PRs introduce a small set of recurring weaknesses such as regex inefficiencies, injection flaws, and path traversal. Many flawed contributions are still merged, while rejections often arise from social or process factors such as inactivity or missing test coverage. The commit message quality of AI PRs has a limited effect on acceptance or latency, in contrast to human PRs reported in previous studies. We also extend existing rejection taxonomies by adding categories that are unique to AI-generated security contributions. These findings offer new insights into the strengths and shortcomings of autonomous coding systems in secure software development.
