Study Finds 38.9% of AI Coding Agent PRs Contain Security Smells

A large-scale empirical study using the AIDev dataset analyzed 16,112 file changes across 4,022 pull requests generated by autonomous coding agents. It found that 38.9% of these PRs contain at least one security code smell, with supply chain integrity issues accounting for 82.3% of all detected smells and hard-coded credentials making up 99.6% of critical-severity findings.

Notably, human collaborators—not the agents themselves—introduced 67.6% of genuine leaked secrets in these agent-assisted workflows. Both automated and human review processes failed to catch 81.1% of these credentials before code was merged, suggesting that developer vigilance may be dropping when working alongside AI agents.

The research highlights that traditional code review is struggling to keep pace with agent-driven development velocity, and calls for context-aware security guardrails built directly into human-AI collaboration points rather than relying solely on post-hoc review.