Google Uncovers First Unattended AI Vulnerability Exploitation Pipeline Where Agents Autonomously Generate and Execute Attack Code
12 hour ago / Read about 0 minute
Author:小编   

Attackers are leveraging AI models to automate the entire process of software vulnerability discovery and exploitation—without any manual, step-by-step intervention. These automated systems can continuously monitor software updates, perform differential analysis of code before and after patches, identify potential vulnerabilities, generate exploit code, and launch attacks automatically. Google's threat intelligence team discovered an exposed, internet-facing server hosting a multi-agent vulnerability research and weaponization framework known as exploit_agent. This framework can connect to multiple types of AI models to analyze common web applications and store exfiltrated data along with various attack scripts. This marks the first confirmed instance of attackers using an AI framework to achieve a fully closed-loop vulnerability discovery and exploitation pipeline. Currently, attackers primarily use AI to rapidly weaponize known vulnerabilities that have already been patched but remain unpatched on a large number of servers—chaining automated differential analysis, dynamic protocol fuzzing, and PoC generation to quickly build exploits after vulnerability disclosure and launch automated attacks against large numbers of servers.