Google Threat Intelligence Group (GTIG) said attackers used large language model (LLM) frameworks to automate software flaw discovery and exploitation without step-by-step human direction. GTIG said the past quarter marked the first time GTIG had identified such activity. Google warned that these automated processes shorten the time organizations have to install security patches before attackers develop exploits.
Researchers detailed the findings in a report released Monday. GTIG found a multi-agent framework for vulnerability research and weaponization on an exposed staging server. The framework continuously compared code versions before and after patch announcements for popular web applications. The tool identified potential flaws, then produced and ran exploit code autonomously. GTIG had said in September that GTIG had not yet seen attackers deploying fully autonomous pipelines against real-world targets.
Attackers are using AI agent frameworks to reverse-engineer patches through differential analysis and develop exploits for N-day vulnerabilities, which are publicly known security flaws. The Google report warned that LLMs allow attackers to expand automated exploitation of internet-facing servers without human intervention. Google said vulnerability disclosures more than doubled between January and August. Google put the monthly average of exploited zero-day vulnerabilities at eight in 2025. Google said the average rose to 11 per month from January through August 2026, with 22 exploited in August.
Automated processes also shorten the time attackers need to progress from initial access to domain administrator privileges. The report cited GreyNoise findings that attackers obtained domain administrator privileges within minutes during an August 2026 PaperCut campaign.
Cloudflare reported attempts in October to exploit PHP flaws against Cloudflare customers. GTIG said LLM analysis had uncovered the two PHP vulnerabilities involved in those attempts. The report said technical details released before the October 21 patch allowed researchers and attackers to examine the vulnerabilities and attempt exploitation.
Anthropic warned in a September report that AI-assisted attacks had developed into multi-agent operations automating most stages of a cyberattack, from discovery through exploitation and data theft. Anthropic said attackers were using agentic AI for campaigns against multiple victims that previously would have needed teams of operators.

