Cybercriminals are increasingly using autonomous artificial intelligence frameworks to identify software vulnerabilities and develop working exploits, reducing the time available for organizations to secure exposed systems.
The November 16 report from Google Threat Intelligence Group documented what its researchers described as their first observations of threat actors automating vulnerability discovery and exploitation without human intervention during individual attack workflows.
The findings represent a significant development in how attackers use generative AI, moving beyond writing malicious code or phishing messages toward coordinating multiple stages of an intrusion.
For software developers and security teams, the immediate concern is not simply more sophisticated attacks, but the speed at which previously disclosed vulnerabilities can become exploitable.
AI Frameworks Automate Vulnerability Discovery
Google researchers Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor identified several offensive AI frameworks, including ARTEX, CAIRN, CyberStrikeAI and PentAGI.
According to the threat intelligence report, these systems can coordinate reconnaissance, vulnerability analysis and exploitation using specialized AI agents.
Google observed CAIRN in intrusion activity targeting organizations in the US aerospace sector.
The company also tracked a customized PentAGI-based platform conducting reconnaissance and vulnerability scanning against organizations in Asia between late August and early October.
CyberStrikeAI, meanwhile, integrates commercial and open-weight language models with more than 100 security tools to automate attack sequences.
These frameworks were originally designed around capabilities also used in legitimate penetration testing, but attackers can adapt them for unauthorized operations.
Attackers Turn Software Updates Into Exploits
One particularly significant discovery involved a framework called exploit_agent.
Google identified the system on an exposed attacker-controlled staging server containing reconnaissance information, scripts and data taken from targeted victims.
The framework compared software code before and after security updates, identified potentially vulnerable changes and attempted to develop and execute exploits automatically.
Investigators found evidence suggesting that the operator targeted French websites in the media, entertainment and education sectors.
This approach focuses on already patched vulnerabilities, often called n-day vulnerabilities, rather than discovering entirely unknown flaws.
Although developers may have released a fix, organizations that have not yet installed it remain exposed.
Vulnerability Research Is Accelerating
Google also cited incidents involving PaperCut print-management software.
In August 2026, researchers and cybersecurity company Huntress used language models to develop working exploits for two disclosed vulnerabilities within one day.
The flaws, identified as CVE-2026-81578 and CVE-2026-82078, involved authentication bypass and unsafe dynamic class loading.
According to Google’s analysis, GreyNoise subsequently observed an exploitation campaign affecting hundreds of organizations, with indications that AI-assisted tools helped accelerate the operation.
Google’s broader tracking also recorded an increase in zero-day exploitation, from a monthly average of eight in 2025 to 11 during 2026, before reaching 22 in August.
These figures describe observed security activity and do not independently establish that AI caused the increase.
Faster Attacks Create New Defensive Priorities
The emerging challenge for engineering teams involves matching the speed of automated exploitation with faster detection, patch deployment and infrastructure monitoring.
Google noted that autonomous offensive frameworks currently sacrifice some stealth for speed, often producing identifiable traffic patterns and forensic traces.
Those limitations may help defenders recognize suspicious activity.
The company has introduced specialized detection rules for ARTEX activity, while Mandiant has published guidance addressing AI-enabled offensive operations.
However, researchers expect attackers to refine their systems as model capabilities improve.
For founders and developers managing internet-facing services, the findings reinforce the importance of rapid vulnerability remediation and continuous monitoring.
The broader lesson is that security teams can no longer assume a lengthy interval between public vulnerability disclosure and attempted exploitation.
As offensive AI becomes more capable, maintaining secure software will increasingly require defensive processes that can respond at comparable speed.