How 360 Trained Tulongfeng to Top the CyberGym Vulnerability Benchmark

What happened

QbitAI reports that 360’s Tulongfeng security system achieved a 99.4% vulnerability reproduction rate on CyberGym, based on three independent evaluations averaging 99.38%. The article describes 360’s approach as combining a security-focused model trained with security data and expert experience with specialized agents that analyze code, use tools, generate PoCs, and verify vulnerabilities. It also recounts a rule the system added after a sandbox blocked its attempt to use an unspecified test address.

Read next

• More changes in today’s briefingGoogle DeepMind Rolls Out Gemini Omni 1.2 Flash for Longer Video Generation

The story so far

360 says its Tulongfeng security system achieved a 99.4% vulnerability reproduction rate on the CyberGym benchmark, the highest score on the leaderboard. QbitAI reports that three independent evaluations averaged 99.38%. The system combines a security-focused model trained on security data and expert experience with specialized agents that analyze code, use tools, generate proof-of-concept exploits, and verify vulnerabilities. QbitAI also describes a rule Tulongfeng added after a sandbox blocked an attempt to use a test address.