Threat Actors Used AI Framework Exploit_agent to Autonomously Weaponize Flaws
Google Threat Intelligence Group reported that attackers used an AI framework called exploit_agent to autonomously discover, develop and execute exploits in real-world campaigns. The finding matters because it shows offensive AI can compress patch-to-exploit timelines and scale n-day attacks across media, education, aerospace and other sectors.
Google Threat Intelligence Group said it discovered exploit_agent on an exposed staging server and assessed that threat actors used it in real-world campaigns to automate vulnerability discovery and exploitation without human intervention. The framework compared patched and unpatched web application code, identified weaknesses, developed exploits and executed them automatically. The server also held stolen victim data and scripts for post-compromise tasks such as reconnaissance and webshell deployment, according to the report.
The evidence trail points to operational automation rather than isolated tool use. Google said exploit_agent was a multi-agent framework built to turn patch differences into working attacks, a task normally done by analysts and exploit developers. The staging server's contents, including victim data and webshell scripts, tied the framework to intrusion activity. Google also linked observed targeting to French media, entertainment and education websites, while attributing the evidence only to target interest, not a named actor.
The report places exploit_agent inside a wider shift toward agentic tooling. Google said attackers use frameworks to reverse engineer patches at machine speed and weaponize n-day vulnerabilities, increasing the velocity of multi-stage intrusions. Artex was the most-used framework, mainly by Chinese-speaking groups, while Cairn, PentAGI, CyberStrikeAI and HexStrike AI were also observed. Cairn appeared in activity targeting US aerospace organizations, and PentAGI in systematic scanning across Asia. Google also described a Chinese group building its own automated attack framework, likely for espionage.
The evidence shows an AI framework can compare patches, write exploits and deploy webshells without an operator, shrinking the window defenders have to prioritize and patch after disclosure. Affected targets include French media, entertainment and education sites, plus US aerospace organizations and Asian companies seen in other framework activity. The risk is a lower-cost path for state-sponsored and cybercrime groups to scale n-day exploitation across sectors.
This affects organizations that patch web applications, especially media, education, aerospace and companies in Asia, because observed AI-driven exploit development can compress the time between patch release and active attacks.
Sources (1)
- [1]AI Framework Autonomously Weaponized Flaws in Real Attacks, Google Says(https://www.securityweek.com/ai-framework-autonomously-weaponized-flaws-in-real-attacks-google-says/)