Google Finds Hackers Using AI Agents to Build and Execute Cyberattacks Without Human Intervention
Share
Hackers are using artificial intelligence agents to automate increasingly complex cyberattacks, including vulnerability discovery, exploit development, and post-compromise activities. Google has identified threat actors deploying multi-agent AI frameworks that can carry out several stages of an intrusion with little human supervision.
The findings come from the Google Threat Intelligence Group (GTIG), which published its latest AI Threat Tracker report on November 16, 2026. The report documents a shift from hackers using AI assistants for individual tasks to deploying autonomous systems that coordinate multiple attack operations.
One of the most concerning discoveries involves an AI-powered framework called exploit_agent. The system can analyze software changes, identify vulnerabilities, develop exploits, and execute them automatically. Google found evidence that the framework had been used against organizations, with data stolen from targeted victims stored on an exposed staging server.
The findings highlight a growing challenge for security teams. Attackers can now use AI to shorten the time between discovering a software weakness and exploiting it, putting organizations under greater pressure to identify vulnerabilities and apply security patches quickly.
Google Identifies AI Agents Automating Multi-Stage Cyberattacks
Google’s investigation found threat actors using several open-source, agentic penetration-testing frameworks to automate different stages of cyber intrusions.
These systems use multiple AI agents that work together to perform reconnaissance, identify weaknesses, generate attack payloads, and adjust their actions based on execution results.
The frameworks identified in Google’s investigation include:
- ARTEX: An AI-powered penetration-testing framework that attackers have used to automate offensive security operations.
- CAIRN: An open-source multi-agent framework that uses an Observe, Orient, Decide, Act (OODA) loop. Its agents coordinate tasks to move from an initial target to an operational objective.
- CyberStrikeAI: An LLM-based offensive security framework that connects AI models with more than 100 security tools, attack recipes, and custom scripts.
- HexStrike AI: An AI-integrated framework that connects language models with more than 150 offensive security tools through specialized agents.
- PentAGI: An autonomous penetration-testing framework that attackers have used for reconnaissance and vulnerability scanning.
Google also observed experimentation with custom-built systems, Agent Zero, and ZeroHollow, a service used for AI inference and exploit development.
The investigation found that financially motivated cybercriminals and suspected China-linked espionage actors were experimenting with these tools.
These frameworks can identify relatively simple vulnerabilities, generate attack payloads when needed, and modify their approach based on real-time feedback. This allows attackers to coordinate multiple operations without manually directing every step.
Google cautioned that these systems currently trade stealth for speed. Their automated activity can generate recognizable network traffic and code-level traces, which security teams may use to identify and investigate intrusions. However, improvements in AI models could make future attacks harder to detect.
AI Framework Automatically Discovers and Exploits Software Vulnerabilities
One of the most important findings in the report concerns exploit_agent, a multi-agent framework that automates vulnerability research and exploitation.
Google discovered the framework on an exposed staging server. It integrates multiple advanced and locally hosted large language models (LLMs) to analyze software continuously.
The framework follows a sequence of automated activities:
- It compares software source code before and after security patches.
- It analyzes code differences to identify potential vulnerabilities.
- It develops exploits based on the weaknesses it discovers.
- It executes the exploits without requiring human intervention.
Google also found evidence of post-compromise activities on the server. These included scripts for internal reconnaissance, brute-force attempts, webshell installation, and removal of webshells belonging to other attackers.
Investigators also found data stolen from targeted organizations.
Victim telemetry suggested that the operator focused on French websites associated with the media, entertainment, and education sectors.
This discovery marks an important development in AI-assisted cybercrime. Google reported that it had identified threat actors fully automating vulnerability discovery and exploitation through LLM frameworks without a human in the loop for the first time.
Hackers Use AI to Turn Security Patches Into Working Exploits
Security patches don’t always end the threat. In some cases, they give attackers valuable information about the vulnerability they need to exploit.
When developers release a patch, they often disclose technical details about the underlying weakness. Attackers can compare the vulnerable and patched versions of the software to understand what changed and identify the flaw.
This process is known as differential analysis.
Google found that threat actors are using LLM-powered systems to automate this process. Instead of manually studying source code, attackers can use AI agents to examine code differences, identify potential weaknesses, and develop proof-of-concept exploits.
The report also highlights attacks involving two PaperCut vulnerabilities disclosed in August 2026:
- CVE-2026-81578: An authentication bypass vulnerability.
- CVE-2026-82078: An unsafe dynamic class loading vulnerability.
The two vulnerabilities could be chained to achieve unauthenticated remote code execution.
Security researchers and Huntress developed working exploits within a day of disclosure. GreyNoise later reported an exploitation campaign in which attackers likely used LLMs to develop and test exploits, identify targets, and attack hundreds of organizations.
In one incident, the attackers reportedly escalated their initial access to domain administrator privileges within minutes.
Google also cited the PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 as examples of AI-assisted vulnerability research and exploitation.
The findings show why organizations cannot assume that releasing a security patch immediately removes the risk. Attackers may use the patch itself to develop an exploit before every affected organization completes the update.
AI-Powered Phishing Platforms Lower the Technical Barrier for Attackers
Google’s investigation also identified the use of AI chatbots in phishing-as-a-service operations.
One example involves a Chinese-language phishing platform called Jieyou. The platform integrates an LLM-powered chatbot that allows customers to customize phishing pages and modify existing templates through natural language instructions.
This reduces the need for technical knowledge when creating phishing campaigns.
Instead of manually editing phishing templates or modifying malicious content, users can communicate with the chatbot to make changes.
The integration of LLMs allows phishing service providers to offer more flexible tools to their customers. It also makes it easier for less technically skilled attackers to create customized phishing campaigns.
For organizations, this creates another challenge. Security teams must look beyond traditional phishing indicators and pay closer attention to how attackers customize messages, impersonate trusted services, and modify phishing infrastructure.
Why AI-Driven Vulnerability Exploitation Demands Faster Patch Management
Google’s findings highlight a growing problem for organizations that rely on conventional vulnerability management processes. Attackers can now use AI agents to analyze software patches, identify the weaknesses they address, and develop working exploits. This reduces the time available for businesses to secure vulnerable systems.
Organizations should take the following steps to reduce their exposure:
- Prioritize critical security patches
Security teams should identify vulnerabilities affecting internet-facing applications, business-critical systems, and widely deployed software. They should prioritize flaws with known exploitation activity and apply vendor patches without unnecessary delays.
- Monitor systems after vulnerability disclosures
Releasing a security patch doesn’t mean attackers will stop targeting the vulnerability. They may reverse-engineer the update to develop exploits. Organizations should monitor affected systems for unusual requests, unauthorized access attempts, and suspicious application behavior.
- Strengthen endpoint and network monitoring
AI-powered attack frameworks can automate reconnaissance, brute-force attempts, and post-compromise activities. Endpoint Detection and Response (EDR) and Extended Detection and Response (XDR) solutions can help security teams identify suspicious processes, privilege escalation, and unusual network connections.
- Review privileged accounts and access permissions
The Google investigation identified post-compromise activities such as brute-force attempts and internal reconnaissance. Businesses should enforce multi-factor authentication, restrict administrative privileges, and regularly review account permissions.
- Prepare for faster incident response
Automated attacks can move quickly from initial access to privilege escalation and data theft. Security teams should maintain incident response plans, define escalation procedures, and use automated containment where appropriate.
The Growing Role of Autonomous AI in Cyberattacks
Google’s findings show that attackers are moving beyond using AI to write malicious code or generate phishing emails. They are beginning to use autonomous agents to coordinate several stages of cyber operations. The discovery of exploit_agent is particularly significant because it demonstrates how AI can automate vulnerability discovery and exploitation without continuous human involvement.
However, the report also identifies limitations. Current autonomous frameworks can leave forensic traces, generate noisy activity, and make mistakes. These weaknesses give defenders opportunities to detect and contain attacks. The challenge is that attackers can improve these systems as AI models become more capable.
For businesses, the priority should be clear: vulnerability management, credential security, continuous monitoring, and incident response must keep pace with the speed of AI-assisted attacks.
Sources
- Google Threat Intelligence Group: AI Threat Tracker, November 16, 2026. Primary research and technical findings.
- The Hacker News: Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches.
- The Register: Google AI Agent Turned a Security Patch Into a Working Exploit Without Human Help.
- Reuters: How One Hacker Used AI Agents to Break Into Scores of Sites.
