AI-generated vulnerability patches require human review

Defenders are increasingly turning to LLMs to to generate vulnerability patches. But our research showed that LLMs only successfully generate patches (without materially changing application behavior) 26% of the time.