Google has identified how AI-powered cyberattacks could take a new step: a framework designed to analyse software patches, find vulnerabilities, draft exploits and execute them without human involvement.
The framework, called exploit_agent, was found on an exposed staging server by Google Threat Intelligence Group (GTIG). In a November 16, 2026 report, Google said this was the first time it had identified threat actors using large language model (LLM) frameworks to automate vulnerability discovery and exploitation without a human in the loop.
Patches can reveal clues about the weaknesses they correct, helping attackers investigate systems that remain unupdated.
How the AI Framework Analyses Patches
Google said exploit_agent connects frontier and locally hosted AI models. It compares web application versions before and after security patches are announced. This process, called patch diffing, helps identify code changes that may reveal the vulnerability a developer has fixed.
The framework analyses weaknesses, drafts exploit code and is configured to execute exploits automatically. Google also found scripts for post-compromise activities, including internal reconnaissance, password brute-forcing and installing webshells providing remote access to compromised websites.
The server contained data taken from targeted victims. Activity suggested interest in French websites in media, entertainment and education. However, Google did not name victims, give a victim count or identify a specific vulnerability successfully exploited by exploit_agent. Its report describes the framework’s capabilities but does not establish that every exploit it drafted worked.
Also Read: OpenAI Resumes Advanced AI Training After Security Overhaul

Why Delayed Security Updates Create Greater Risks
A known vulnerability with an available fix is called an n-day vulnerability. Attackers can study the patch to understand the weakness. AI can reduce the time and effort needed to conduct this research and test possible attacks.
That does not make every AI-generated exploit successful. In a separate PHP case, Google found hundreds of claimed but non-functional proof-of-concept exploits, likely AI-generated. An attack still requires an exposed system and an exploit that works. Automated analysis can shorten the time organisations have to secure exposed systems.
Organisations should track internet-facing applications, plugins, servers and third-party components, then prioritise updates according to security notices. Patching should be an assigned responsibility, not occasional maintenance.
What Nigerian Businesses and Institutions Should Do
The findings matter to Nigerian banks, schools, media organisations and small businesses that depend on websites and online services. Automated attackers can search for vulnerable systems without knowing an organisation beforehand.
Website owners should update content management systems, themes, plugins and server software, remove unused components, enable multi-factor authentication and keep tested backups. They should also list internet-facing systems and assign responsibility for security updates.
Smaller businesses without in-house security teams should agree on patching and incident-response responsibilities with website developers or IT providers. A website functioning normally does not mean its software is secure.
Back Story: Google’s Earlier Warning on AI-Enabled Attacks
In a September 8, 2026 report, Google described attackers moving from basic AI prompts to multi-agent workflows. It documented a campaign in which an AI-assisted process helped plan and execute mass credential harvesting in less than six hours.
The report marks a further step: a framework configured to automate vulnerability research and exploitation in one pipeline. Patches are not inherently dangerous; the risk is leaving known weaknesses unfixed.
Also Read: Hackers Turn to AI to Automate Cyberattacks Against African Educational Institutions


