Google Identifies Autonomous AI Framework Used by Hackers to Exploit Software Flaws
Share
Researchers from the Google Threat Intelligence Group (GTIG) and Mandiant have identified a hacker-operated AI framework designed to find and exploit software vulnerabilities autonomously. Known as “exploit_agent”, the multi-agent system marks a significant evolution from AI-assisted tasks to fully automated cyberattack chains.
The framework was discovered on an exposed staging server and is configured to connect to several large language models (LLMs). Its primary function is to analyse security patches and application code to identify vulnerabilities. By comparing version changes, the system can autonomously generate and execute exploits for “n-day” vulnerabilities—flaws that have been publicly disclosed but not yet patched by all organisations.
Automated Exploitation in the Wild
During their investigation, researchers found evidence of data exfiltration on the server, suggesting the system or its operators had already successfully targeted victims. Specific interest was observed in French websites within the education, media, and entertainment sectors. While researchers have not yet confirmed if the autonomous features were solely responsible for the data theft, the presence of post-compromise scripts indicates a high level of operational maturity.
Google also highlighted other AI-driven frameworks used in active campaigns. A framework named CAIRN was identified in attacks against the US aerospace sector, while another platform based on PentAGI was used for reconnaissance and scanning across Asia. In a separate instance, threat actors appeared to use AI tools to rapidly develop exploits for PaperCut print management servers, launching attacks against hundreds of organisations just four days after the vulnerabilities were disclosed.
Pressure on Enterprise Response Times
Security analysts suggest that while AI significantly accelerates the speed of attacks, the underlying techniques remain largely consistent with established red-teaming and penetration testing methodologies. The primary risk shift is the compression of the response window for security teams. The ability of AI to rapidly digest patch notes and generate exploit code means the “race to patch” has become significantly more difficult for defenders.
Despite the speed of these frameworks, they typically generate a high volume of detectable activity. Experts note that properly configured endpoint detection and response (EDR) and managed detection and response (MDR) systems should still be capable of identifying the malicious behaviour. However, the speed of attacker movement requires a corresponding increase in the speed of containment.
Security leaders are advised to prioritise the hardening of cloud and identity systems. Organisations should treat internet-facing systems as a priority class, implementing fast-track patching paths or virtual patching while official fixes are tested. Furthermore, experts recommend automating well-understood containment actions, such as session revocation or endpoint isolation, to eliminate avoidable delays in the incident response process.
The discovery of exploit_agent suggests that open-source agentic frameworks are lowering the technical barriers for less-experienced criminals, allowing them to coordinate complex tools and adapt attacks based on real-time results. For now, the most effective defence remains the rigorous application of established security practices, accelerated to match the pace of AI-driven threats.




Leave a Reply