CyberGym, a benchmark developed by researchers at UC Berkeley, reveals that the best artificial intelligence agents now achieve a 93.2% success rate in autonomously reproducing real-world security vulnerabilities, compared to just 10-30% a year ago. The leaderboard is topped by the Sangfor AI Agent running on DeepSeek-V4-Flash, followed by three other agents all exceeding the 90% threshold. The evaluation draws from 1,507 real vulnerability instances across 188 open-source projects, primarily sourced from Google’s OSS-Fuzz corpus. Beyond reproducing known bugs, the agents have collectively discovered 34 new zero-day vulnerabilities and identified 18 incomplete patches, demonstrating a systematic capability to verify the reliability of security fixes.
Source: Read the original article
