Google Threat Intelligence Group and Mandiant revealed in a November 16 report that attackers were using AI frameworks to identify vulnerabilities, generate malicious code, and execute exploits with minimal human involvement.
Researchers discovered a system called exploit_agent on an exposed staging server containing multiple AI models connected to cybersecurity research and exploitation tools. The framework monitored software updates, compared patched code with older versions, and identified weaknesses that developers had recently repaired. It could subsequently generate attack code and execute it automatically, allowing attackers to target websites that had not yet installed security updates.
This approach exploits information developers publish when addressing software vulnerabilities. Changes made to repair a flaw can inadvertently reveal how the underlying weakness could be exploited, allowing attackers to focus on known vulnerabilities rather than spend time discovering previously unknown security flaws.
Google's investigation also uncovered stolen victim data and malicious scripts capable of conducting internal network reconnaissance, brute-force attacks, and installing webshells that provide remote access to compromised servers. Some tools could remove webshells left by competing attackers, potentially helping operators retain control over compromised systems.
Google found evidence that one operator was particularly interested in French websites associated with media, entertainment, and education, including examination of access-control protections in Spectra, a WordPress plugin used for website development.
Cryptocurrency Firms Targeted Through AI-Generated Recruitment Fraud
Google's report also revealed a separate campaign in which suspected North Korean hackers exploited AI-powered development tools to infiltrate cryptocurrency organizations in North America and Europe.
The attackers, tracked as MIDNIGHT NEPTUNE, combined AI-generated identities with fraudulent recruitment campaigns. They created fake recruiter profiles on professional networking platforms using AI-generated photographs, then directed prospective victims to fraudulent meeting websites that instructed them to run supposed troubleshooting commands that instead executed malicious scripts.
Once inside an organization, the attackers accessed internal code repositories and modified configuration files used by Anthropic's Claude Code development assistant. The altered files were designed to automatically download a backdoor known as SOMBERMEME whenever employees launched Claude Code. The attackers also disguised GitHub access tokens using names associated with Microsoft Copilot to conceal their movements within compromised development environments.
Implications for Software Security
Google warned that automating vulnerability analysis and exploit execution could substantially reduce the time between disclosing a security flaw and exploiting it, leaving website operators with increasingly narrow windows to install protective updates. The discovery marked the first time Google had identified attackers using AI frameworks to fully automate vulnerability discovery and exploitation without human intervention throughout the process.


