Hackers use AI to automate vulnerability research, and exploited zero-day vulnerabilities surge to 22 in a single month
駭客利用AI自動化漏洞研究,遭到利用的零時差漏洞一個月爆增至22個
Google Threat Intelligence Group (GTIG) released an AI threat tracking report on November 17, stating that hackers' use of AI agents and LLMs for vulnerability research and attack tool development has moved from the experimental stage to actual operations, and that exploited zero-day vulnerabilities increased from a monthly average of 8 in 2025 and a monthly average of 11 from January to July this year to 22 in August. Attackers use multi-agent systems to perform patch diffing, dynamic fuzzing and PoC exploit development, and can also use LLMs to quickly grasp the root cause of known vulnerabilities with public information and develop exploits; among them, the PaperCut vulnerability exploited starting in late August and the PHP critical vulnerabilities collectively called PHP2Shell both saw exploitation activity within days after PoC information or related patch code was released. GTIG also found an exploit_agent multi-agent vulnerability research and weaponization framework running on an internet-exposed test server, which can automatically compare code before and after a software security patch is released, identify the vulnerability, and generate and execute an exploit; it is the first fully human-free vulnerability discovery and exploitation platform observed by researchers, and the victim organization data it stored shows that the targets were concentrated on websites in the French media, entertainment and education sectors.
The report provides monthly statistics on exploited zero-day vulnerabilities and the fully automated exploit framework first observed by GTIG, which defenders can use to assess the response window after a patch is released.