The framework, named VEXAIoT, uses cooperating AI agents to scan networks, identify vulnerable services, devise attack plans and execute exploits with limited human involvement. Researchers tested it against IoTGoat and Metasploitable2, two deliberately vulnerable platforms used for cybersecurity training and controlled experimentation.
VEXAIoT achieved a 94.5% success rate on IoTGoat, completing 189 of 200 attack attempts. It succeeded in 58 of 60 trials against Metasploitable2, producing a rate of 96.7%. The combined result was 247 successful executions from 260 attempts.
The findings demonstrate how large language models can move beyond offering security advice or generating code to coordinating complete penetration-testing workflows. Most successful attacks were executed in under two minutes, while several scenarios produced success rates of 100%.
The study was conducted by Katherine Swinea, Kshitiz Aryal, Lopamudra Praharaj and Maanak Gupta. Their research was submitted on July 10 and has not yet undergone formal peer review, meaning its findings require independent validation before they can be treated as evidence of performance in operational networks.
VEXAIoT separates its work between two specialised components. A vulnerability detection agent performs reconnaissance and collects details about available hosts, ports, services and software versions. An attack execution agent analyses those findings, selects a possible exploitation route and invokes offensive security tools.
This division allows the system to revise its strategy when a command fails rather than following a fixed script. Large language model reasoning is used to interpret tool outputs, connect discovered services with likely vulnerabilities and decide which action to take next.
The researchers evaluated ten attack scenarios aligned with categories in the Open Worldwide Application Security Project’s Internet of Things security framework. These included weak or hard-coded passwords, exposed network services, insecure configurations and software components carrying known weaknesses.
IoTGoat, maintained as an intentionally insecure firmware project, is designed to reproduce common flaws in connected devices. Metasploitable2 provides a wider collection of vulnerable network services. Both systems operate as training targets and do not represent the defensive controls, device diversity or unpredictable conditions found in production environments.
That distinction limits the broader meaning of the 95% figure. The result does not show that VEXAIoT could compromise 95% of consumer devices, industrial equipment or smart-home products. It measures repeated attempts against predefined vulnerabilities in controlled test systems where exploitable weaknesses were deliberately present.
Real networks may contain patched software, intrusion detection systems, segmented architectures, access restrictions and incomplete technical information. Connected devices also use varied processors, operating systems and proprietary protocols, all of which can complicate automated exploitation.
Even so, the experiments highlight a shift in offensive security automation. Conventional vulnerability scanners can identify exposed services and match software versions with databases of known flaws. They usually require a human tester to interpret findings, select an exploit and adapt when the first approach fails.
Agent-based systems can potentially combine those stages. They can read scan results, reason about possible attack paths, issue commands and evaluate whether an attempted compromise worked. This could help authorised security teams examine large device fleets faster and prioritise weaknesses before they are abused.
The same capability carries significant risks. Systems that lower the expertise needed to conduct multi-stage attacks could be misused against poorly maintained cameras, routers, sensors and industrial controllers. IoT products often remain operational for years without firmware upgrades and may be deployed with default credentials or unnecessary services exposed to the internet.
The researchers confined their work to isolated test environments and positioned the framework as a tool for authorised vulnerability assessment. Safe deployment would require strict access controls, comprehensive logging, human approval for sensitive actions and rules preventing agents from reaching systems outside an approved testing scope.
VEXAIoT also relies on established offensive tools rather than discovering entirely new exploitation techniques. Its main contribution is the autonomous coordination of reconnaissance, reasoning and attack execution. The study reports low language-model token consumption, suggesting that repeated tests may be conducted without the high computational expense associated with long conversational prompts.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
