99.4% and World No. 1: This Chinese AI Learned from World Champions
The self-developed AI security system Tulongfeng from 360 has topped the CyberGym leaderboard with a 99.4% vulnerability-reproduction success rate, ahead of GPT-5.5-Cyber from OpenAI, Claude Mythos Preview from Anthropic and Gemini 3.8 Flash Cyber from Google. Across 1,507 real vulnerability tasks it scored 99.27%, 99.47% and 99.40% in three independent runs, and since its June launch it has found more than 15,000 vulnerabilities. It matters because it shows AI agents moving from spotting known bugs to reliably reproducing real ones — a major step for automated security — and a Chinese system beating frontier US models on a public benchmark.
SOURCES
- 量子位 QbitAI2026-11-06 05:01