New AI tool finds 85.7% of bugs in AI-generated code, cuts testing time by 1.71x

Researchers have unveiled SAFuzz, a new AI-powered fuzzing framework designed specifically to test code generated by LLMs. It uses semantic guidance and adaptive resource allocation to find algorithmic vulnerabilities more efficiently. The system improves vulnerability detection precision from 77.9% to 85.7% and reduces time costs by 1.71x compared to the current state-of-the-art, GreenFuzz. When combined with existing unit test methods, bug detection recall jumps from 67.3% to 79.5%.