GTIG, Google's threat intelligence division, and security firm Mandiant have confirmed the first actual cyberattack powered entirely by AI . Unlike previous instances where hackers used chatbots to assist in writing malicious code, this system handles all stages from vulnerability detection to exploitation itself.

Specifically, Google investigators discovered a test server on the network running a multi-agent system called exploit_agent, connected to several large, locally installed language models to evaluate software updates.

Exploit_agent uses differential patch analysis, comparing source code before and after a security patch is released to identify newly fixed vulnerabilities. From there, the AI ​​automatically writes exploit code and launches attacks on systems that haven't been updated with the patch, a process that occurs without human intervention.

Logo Google tại trụ sở hãng ở Mountain View, California, Mỹ, tháng 5/2026. Ảnh: Lưu Quý
The Google logo at the company's headquarters in Mountain View, California, USA, May 2026. Photo: Luu Quy

In addition to vulnerability hunting capabilities, this server also contains numerous post-intrusion scenarios such as scanning the internal network, automated password cracking, and installing remote malware to maintain access.

Data collected indicates that the targets were primarily media, educational , and entertainment websites in France. Google has not attributed responsibility to any specific hacking group or country.

According to a Google report, the use of multi-agent AI systems is becoming a new trend among cybercriminals. Instead of relying on a single tool, hackers are splitting the process among specialized AIs: one agent researches the target, another detects vulnerabilities, and a third writes and tests exploit code before deciding the next step.

Google has named a number of open-source attack tools being tested by hackers, including ARTEX, CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI. One group, in particular, customized the PentAGI framework to scan for vulnerabilities in numerous organizations across Asia from late August to early October.

According to Tom's Hardware , AI support has nearly doubled the number of vulnerabilities disclosed and exploited from January to August. Attacks exploiting zero-day vulnerabilities (those without patches) alone reached 22 in August, double the average of 11 per month in 2026 and a significant increase from the 8 per month in 2025.

Although AI-written malware still has a high error rate and its automated operation often leaves more easily detectable network traffic trails, security experts believe the cybersecurity landscape is changing rapidly. Google advises organizations to minimize patch update times, tightly protect cloud accounts, and promptly equip themselves with AI-powered automated defense systems to counter the new pace of attacks.