Google Threat Intelligence Group and Mandiant have identified a significant shift in the cybersecurity landscape, confirming that threat actors are now utilizing autonomous AI agents to execute full-scale cyber intrusions without human intervention. This development marks the first time researchers have observed LLM-powered frameworks capable of independently identifying vulnerabilities, drafting exploits, and launching attacks in real-time.
Evidence of this capability was discovered on an exposed staging server hosting a multi-agent framework identified as exploit_agent. This system was configured to monitor web application repositories, compare code before and after security patch releases, identify flaws, and immediately deploy functional exploits. The server also contained scripts for brute-force attacks and the management of webshells. Configuration files associated with the framework, which included Russian-language comments, explicitly labeled the integrated LLM API keys as being intended for automated exploitation.
Beyond fully autonomous agents, threat actors are increasingly adopting open-source agentic pentesting frameworks to accelerate their operations. The FBI issued a warning on November 13 regarding the use of the ARTEX framework, which has been deployed against organizations in the defense, government, financial, and healthcare sectors since August 2026. Other tools identified in active campaigns include CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI. While these frameworks currently prioritize speed over stealth, researchers note that they are being utilized by both financially motivated cybercriminals and suspected state-sponsored groups, including those with ties to the People's Republic of China.
The impact of these tools is reflected in a sharp rise in vulnerability exploitation. Google’s analysis indicates that the number of zero-day exploits surged to 22 in August 2026, compared to a 2025 monthly average of eight. Furthermore, defenders are facing a deluge of non-functional, AI-generated proof-of-concept code, which complicates incident response efforts. For instance, following the release of patches for PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 in October, hundreds of fake exploits flooded the internet, requiring significant manual verification by the security community.
Mandiant has also reported a rise in LLMJacking, a technique where attackers compromise cloud environments to hijack and exhaust a victim’s LLM compute quotas. These incidents have resulted in fraudulent charges exceeding $100,000 in single-day bursts, with some cases reaching $1 million in total damages. Stolen access to these models is frequently resold on underground platforms referred to as LLM Mixers, where frontier and jailbroken models are available for purchase at low costs.
North Korean threat actors have also integrated AI into their operational workflows, particularly through the use of fake personas. A cluster identified as MIDNIGHT NEPTUNE has been observed using AI-generated recruiter profiles—verified by the presence of SynthID watermarks—to target cryptocurrency organizations. These actors utilize ClickFix lures to trick victims into running malicious scripts. Once inside a network, they have been observed modifying configuration files, such as .claude/settings.json, to force the automatic download of a backdoor known as SOMBERMEME whenever legitimate development tools are executed.
Finally, influence operations have moved toward full automation. Researchers investigated a platform designed to manage over 100,000 social media accounts, which utilized LLMs to simulate human behavior, such as scrolling and browsing, to "age" accounts before deploying political messaging. These systems are programmed to mimic local dialects and slang, and they intentionally introduce typos and image artifacts to bypass automated detection systems.
While these automated frameworks currently leave behind identifiable code and traffic traces, researchers warn that the trend toward machine-speed intrusions and the automated chaining of complex attack paths represents a fundamental shift in the threat environment. As LLM sophistication grows, the expectation is that these operations will become increasingly stealthy, further reducing the time available for human defenders to respond.





