Google says it found a hacker AI framework that compares patches and automatically generates attack code
补丁成了「漏洞说明书」!谷歌曝黑客让AI照着补丁写攻击代码
Google Threat Intelligence Group (GTIG) says it found an AI attack framework named exploit_agent on a hacker staging server directly exposed to the public internet; it continuously compares code differences before and after software patches to infer vulnerability locations, then automatically generates and executes attack code. Google says this is the first time it has found attackers using a large-model framework to string together the entire process from finding vulnerabilities to exploiting them with no human involvement in between. The framework connects to multiple frontier models and local models at the same time and targets popular web applications; the reading record Google showed was a piece of access control logic in the WordPress plugin Spectra 2.19.25. Google did not disclose the targets it actually compromised; its configuration file still requires a human to enter an LLM API key, and Google describes such frameworks as trading stealth for speed and being noisy and easy to expose. The figures given in the report are that vulnerabilities actually used in attacks rose from a monthly average of 10.5 in 2025 to nearly 18 in the first 8 months of 2026, among which zero-day vulnerabilities rose from a monthly average of 8 to 11, reaching 22 in August alone.