Google identifies first fully autonomous AI cyberattack framework
Google researchers have identified the first fully autonomous AI-powered cyberattack framework capable of discovering and exploiting vulnerabilities without human intervention. The Google Threat Intelligence Group (GTIG) published these findings in its Q4 2026 AI Threat Tracker on November 16, 2026. This discovery marks a significant shift from using artificial intelligence to accelerate human-led attacks to AI performing the entire exploitation pipeline autonomously.
Autonomous exploitation framework details
Researchers identified an autonomous framework named exploit_agent hosted on an exposed staging server. This system utilizes multiple AI agents alongside both commercial and local language models to compare software code before and after security patches are released. The framework is capable of automatically identifying vulnerabilities, generating exploit code, and executing attacks without human oversight. According to Google, the era of autonomous attacks has arrived.
The identified attacks targeted various French websites within the media, entertainment, and education sectors. During the investigation, researchers found Russian-language comments within the system's configuration files, which suggests the presence of a Russian-speaking operator.
Rising trends in zero-day attacks
The report highlights a significant increase in the volume of cyber threats. The number of vulnerabilities and exploits disclosed between January and August 2026 nearly doubled compared to previous periods. Specifically, zero-day attacks targeting unpatched vulnerabilities have seen a steady rise. These attacks rose from a monthly average of 8 in 2025 to an average of 11 in 2026. The data showed a significant spike in August 2026, when 22 zero-day attacks were recorded.
In addition to autonomous frameworks, researchers documented sophisticated "LLMJacking" incidents. One such incident handled by Mandiant resulted in over $100,000 in unauthorized usage fees in a single day. The most severe case of LLMJacking resulted in a total cost of $1,000,000.
Threat actor activity and tactics
The UNC6836 group, which is linked to North Korea, has utilized AI-generated fake recruiter profiles to target crypto-asset organizations in North America and Europe. These profiles contained SynthID watermarks. Furthermore, the UNC6836 group modified configuration files for Anthropic's Claude Code to download backdoor malware when developers initiated the tool.
Other operators are using large-scale automation to manage disinformation and social engineering. One operator utilized a platform to manage over 100,000 accounts. This operator used large language models to mimic human behaviors, such as scrolling and reading, specifically to avoid detection by security systems.
Glossary
- zero-day attacks — These are attacks targeting unpatched vulnerabilities. They have seen a steady rise in volume.
- LLMJacking — LLMjacking refers to incidents involving unauthorized usage fees. These incidents are described as sophisticated.
- Claude Code — Claude Code is an agentic command line tool used to delegate coding tasks from the terminal using natural language prompts. It is an agentic tool made by Anthropic.
- social engineering — Social engineering is a tactic used by operators. It can be managed through large-scale automation.
- security patches — Security patches are used to fix security vulnerabilities in software. They are compared by autonomous frameworks before they are released.
This article was written with AI assistance from multiple news reports and is reviewed by our editors under our editorial principles. Editorial Standards · Tips & Contact
Comments 0