Google’s Threat Intelligence Group and its Mandiant incident response unit have identified what they characterize as the first fully automated cyberattack conducted without human intervention. According to the company, the operation represents a significant shift in the threat landscape, moving away from AI-assisted tasks—such as drafting phishing emails or summarizing stolen data—toward an autonomous pipeline that manages the entire lifecycle of an intrusion, from initial vulnerability research to final exploitation.
The researchers identified a specific system, which they have named exploit_agent, as the centerpiece of this operation. This framework utilizes a combination of frontier and locally hosted language models to monitor software updates. By comparing code before and after a vendor releases a patch, the system is designed to infer the nature of the vulnerability, draft a functional exploit, and execute it automatically. The system also includes scripts capable of performing internal reconnaissance, guessing passwords, and deploying webshells.
Google’s analysis suggests the attackers are Russian-speaking, a conclusion based on the presence of Russian-language comments within the operation's source code. One configuration file explicitly instructed users to input API keys for auto-exploitation directly into the system rather than through a chat interface. While the researchers have linked the activity to Russian-speaking actors, they have not formally attributed the campaign to a specific state-sponsored group or known criminal organization.
The targeting data recovered from a staging server indicates that the automated campaign was focused on French websites, specifically those within the media, entertainment, and education sectors. Beyond the primary exploit_agent, investigators observed the system performing secondary tasks, such as erasing the webshells of rival attackers who were operating on the same compromised infrastructure.
The report also highlights a broader trend of autonomous offensive frameworks emerging in the wild. Google identified the use of open-source tools such as CAIRN and PentAGI in separate campaigns targeting an aerospace organization in the United States and various entities across Asia. Additionally, the company noted that a China-nexus group is currently developing its own automated framework, which analysts believe is intended for espionage purposes.
The shift toward automation has significantly increased the speed at which vulnerabilities are weaponized. Following the disclosure of two PHP vulnerabilities on October 21, hundreds of purported exploits emerged within days. While many of these were nonfunctional and likely generated by less sophisticated AI tools, the trend underscores the growing pressure on defenders. Google reported that the number of exploited zero-day vulnerabilities reached 22 in August 2026, a sharp increase from the 11 per month observed earlier that year and the 2025 average of eight.
Google characterizes this development as a critical inflection point in cybersecurity. The ease of integrating new exploits, combined with the ability to scale operations rapidly, suggests that autonomous attack frameworks will likely become a standard tool for threat actors globally. The company warns that the low barrier to entry for these technologies will continue to drive the adoption of fully automated intrusion pipelines.
These findings coincide with a period of heightened scrutiny regarding the security implications of agentic AI. The report references earlier incidents, including a summer event where an OpenAI agent successfully compromised the development platform Hugging Face, as well as the release of cybersecurity-focused models like Anthropic’s Mythos. As these systems become more capable, the challenge for incident responders is expected to grow, as the speed of automated exploitation often outpaces traditional human-led defensive measures.





