A new multi-agent framework called VEXAIoT uses large language model (LLM) reasoning to autonomously discover and exploit vulnerabilities in Internet of Things (IoT) environments.
Developed by Katherine Swinea and colleagues at Tennessee Tech University, the system achieved a 95.0% overall success rate across 260 attack executions.
IoT devices remain a persistent security weak point because of constrained hardware, outdated firmware, and insecure default configurations, which creates demand for scalable, adaptive testing methods.
While LLM agents have already shown promise in general penetration testing and Capture-the-Flag competitions, their application to IoT-specific attack surfaces had remained largely unexplored until this work.
VEXAIoT AI Agents Achieve 95% Success
VEXAIoT is built around two coordinated AI agents: a vulnerability detection agent and an attack execution agent.
The detection agent scans the target using nmap, cross-references discovered services against Searchsploit’s exploit database, and feeds this reconnaissance data into an LLM (ChatGPT 5.1 Thinking) to plan an attack sequence.
That plan is then handed to the attack execution agent, which selects appropriate tools or scripts such as BetterCap or John the Ripper, generates executable commands, and runs them against the target device.
Results, including successes, failures, and any error messages, are looped back to the detection agent, which validates outcomes and either proceeds to the next attack or triggers a retry with an adjusted strategy.
This closed-loop feedback design lets the system adapt in real time rather than following a static, predefined script. The framework was evaluated in two intentionally vulnerable environments: IoTGoat, an OpenWrt-based router firmware, and Metasploitable2.
The IoTGoat testbed used a three-machine setup: a Kali Linux attacker host running the agents, an Ubuntu 24.04.4 client, and the IoTGoat router as the target connected to simulate realistic network traffic flows.
Ten attack scenarios were mapped directly to nine of the ten OWASP IoT Top 10 categories, covering weak passwords, insecure network services, insecure ecosystem interfaces, insecure update mechanisms, outdated components, insufficient privacy protection, insecure data transfer, lack of device management, and insecure default settings.
Physical hardening was excluded because it falls outside the framework’s network- and firmware-level scope.
Each of the 10 IoTGoat scenarios was run 20 times, totaling 200 trials. The vulnerability detection agent successfully completed reconnaissance and planning in all 200 cases, meaning every failure occurred during exploit execution, not discovery.
The attack execution agent completed 189 of those 200 trials successfully, a 94.5% success rate, with seven attacks, including XSS, developer backdoor access, malicious update triggering, PII extraction, log erasure, and remote code execution, reaching a perfect 100% success rate.
In the separate Metasploitable2 environment, the framework achieved a 96.7% success rate, contributing to the combined 95.0% figure across both testbeds.
The weakest performers were the MiniUPnP backdoor and DNS denial-of-service attacks, each achieving 80% success, largely due to their reliance on precise command syntax and service-specific interactions that proved more error-sensitive.
A man-in-the-middle credential-grab attack achieved 95% success once the network prerequisites were established.
Failure analysis attributed most shortfalls to persistent syntax-generation errors (5 instances) and model refusals (5 instances), with hallucinated outputs accounting for just one failure.
By separating reconnaissance/planning from execution while keeping both agents coordinated through shared context, VEXAIoT demonstrates that LLM-driven agents can automate substantial portions of IoT penetration testing workflows in controlled lab environments.
The Arxiv notes that while VEXAIoT demonstrates strong feasibility for automating IoT security testing in controlled environments, reliable autonomous exploitation still demands stronger command validation, safety controls, error recovery, and defensive agents capable of mitigating discovered vulnerabilities.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
 reasoning to autonomously discover and exploit vulnerabilities in Internet of Things (IoT) environments.){kind=link}