VEXAIoT Multi-Agent System Automates IoT Reconnaissance and Exploit Execution

Key Takeaways

• Researchers have unveiled VEXAIoT, an AI-driven multi-agent system designed to automate the discovery and exploitation of vulnerabilities in Internet of Things (IoT) devices.
• The system achieved a 95% success rate across 260 attack attempts in controlled, intentionally vulnerable environments like IoTGoat and Metasploitable2.
• VEXAIoT coordinates reconnaissance, attack planning, command generation, and result validation using two distinct AI agents.
• While demonstrating significant potential for penetration testing, researchers caution about challenges such as AI “hallucinations” and the critical need for human oversight.

Cybersecurity researchers have developed VEXAIoT, an innovative AI-powered multi-agent framework engineered to streamline the process of identifying vulnerabilities and executing exploits within Internet of Things (IoT) ecosystems.

This groundbreaking research demonstrates how large language model agents can effectively collaborate to perform reconnaissance, formulate attack strategies, generate precise commands, and validate outcomes within isolated security testbeds.

VEXAIoT: An AI-Driven Attack Framework

VEXAIoT, an acronym for Vulnerability EXploitation using AI Agents, operates through the coordinated actions of two interconnected agents. The first, a vulnerability detection agent, is tasked with scanning target devices, mapping exposed services, and identifying known vulnerabilities.

The second component, an attack execution agent, is responsible for selecting appropriate tools, crafting the necessary commands, and attempting the planned exploit. The detection agent initiates its process by employing Nmap to pinpoint open ports, active services, and network protocols in use.

Following this initial scan, it cross-references the discovered software and version information with data from Searchsploit and the Exploit Database. This step aims to match identified software to known Common Vulnerabilities and Exposures (CVEs) and publicly available proof-of-concept exploits.

Subsequently, the AI model analyzes this collected intelligence to construct a prioritized attack plan. This plan considers factors such as vulnerability severity, the availability of exploitation tools, and any interdependencies between potential attacks.

Automating IoT Reconnaissance and Exploitation

The VEXAIoT framework exhibits advanced capabilities, such as adapting its approach based on attack prerequisites. For instance, if an exploit requires valid credentials, VEXAIoT will first attempt credential recovery or network traffic interception before proceeding with the dependent actions.

Furthermore, the system is designed to reattempt failed attacks. It leverages error messages and execution outputs to refine its commands and adjust its subsequent attempts, demonstrating a degree of autonomous learning and adaptation.

Researchers rigorously evaluated VEXAIoT against two intentionally vulnerable environments: IoTGoat, an OpenWrt-based IoT firmware specifically designed for security testing, and the well-known Metasploitable2 vulnerable machine.

The IoTGoat evaluations encompassed ten distinct scenarios aligned with OWASP IoT security risks. These included common issues such as weak passwords, insecure network services, exposed developer backdoors, insecure update mechanisms, DNS denial-of-service, plaintext sensitive data exposure, man-in-the-middle interception, remote code execution, and log deletion.

Across 200 attack attempts on IoTGoat, VEXAIoT successfully completed 189, achieving an impressive 94.5% success rate. Seven of these scenarios, including cross-site scripting, developer-backdoor access, malicious update execution, database PII extraction, log erasure, and remote code execution, yielded a perfect 100% success rate.

The lowest success rates, at 80%, were observed in tests involving the MiniUPnP backdoor and DNS denial-of-service, primarily due to command syntax issues and instances where the AI model refused to proceed.

According to arXiv, the framework also demonstrated strong performance against Metasploitable2, successfully exploiting the VSFTPD backdoor and uncovering exposed database credentials in all 60 tests. It achieved remote code execution in 18 out of 20 attempts.

Cumulatively, VEXAIoT achieved a 95% success rate across 260 executions. Most attacks were completed in under two minutes, though password cracking operations naturally required more time.

The research also highlighted efficiency gains from parallelizing independent attacks, reducing the total test time from approximately eight minutes and 31 seconds to about three minutes and 50 seconds. However, token consumption remained largely consistent, as each attack still necessitated a separate interaction with the AI agent.

This study underscores the increasing potential of agentic AI in authorized IoT penetration testing and vulnerability validation. Nevertheless, the authors caution that autonomous exploit execution still faces significant hurdles, including the generation of “hallucinated” or incorrect outputs, invalid commands, model refusals, and the paramount need for robust human oversight and safety protocols.

It is crucial to note that this framework was evaluated exclusively within controlled, intentionally vulnerable environments and must not be deployed against any systems without explicit and authorized consent.

What You Should Do

• Regularly update IoT device firmware and software to the latest versions.
• Implement strong, unique passwords for all IoT devices and services.
• Segment IoT devices onto a separate network to limit their exposure and potential impact in case of compromise.
• Disable unnecessary services and ports on IoT devices.
• Conduct regular vulnerability assessments and penetration tests, utilizing authorized tools and methodologies.

Disclaimer: HackersRadar reports on cybersecurity threats and incidents for informational and awareness purposes only. We do not engage in hacking activities, data exfiltration, or the hosting or distribution of stolen or leaked information. All content is based on publicly available sources.