AI Agents Hacked IoT Systems in 95% of Tests—Autonomously
Researchers built a two-agent AI system that autonomously found and exploited IoT vulnerabilities, succeeding in 95% of 260 attack runs across two test environments.
In Brief
- A two-agent AI system autonomously found and exploited IoT vulnerabilities, succeeding in 95% of 260 attack runs.
- The framework, VEXAIoT, pairs a vulnerability-detection agent with an attack-execution agent across IoTGoat and Metasploitable2.
- Most successful attacks finished in under two minutes, with low token overhead in the tested environments.
Researchers introduced a multi-agent AI system that autonomously hunts for and exploits vulnerabilities in Internet-of-Things devices—with no human at the keyboard. Across 260 attack executions, it succeeded 95% of the time.
The framework, called VEXAIoT, combines a vulnerability-detection agent and an attack-execution agent that perform reconnaissance, plan attack sequences and run exploits against simulated vulnerable IoT services. The work appears on the arXiv preprint server.
Most attacks wrapped up in under two minutes. The system was tested in IoTGoat and Metasploitable2, two environments mapped to OWASP’s catalog of IoT weaknesses, spanning ten attack scenarios.
How the Autonomous AI Agents Work
VEXAIoT splits the job between two specialized agents. One scouts a target and flags likely weaknesses; the other turns those findings into working exploits using offensive-security tooling, the researchers wrote in the paper.
The division paid off. The system reached a 94.5% success rate against IoTGoat and 96.7% against Metasploitable2, with individual attack success reaching as high as 100% in some scenarios. The authors note the runs used low token budgets.
The results show LLM-driven agents can automate IoT vulnerability assessment and offensive workflows in controlled environments, the paper states. The authors stress the tests were carried out in sandboxes, not against live infrastructure.
Why Autonomous Exploitation Matters
The finding lands as defenders already struggle to patch IoT gear with outdated firmware and insecure default settings. An AI agent previously found an unpatched remote-code flaw in roughly 30% of the web, showing the same class of tooling works on both sides.
Security agencies have pushed hard on remediation. CISA recently gave federal teams three days to patch a Check Point VPN bug already exploited by ransomware crews, a sign of how thin the margin for automatic exploitation has become.
For now the work is a benchmark, not a deployed weapon. But it sketches a near future where the gap between finding a flaw and firing it shrinks from days to seconds—and where the fastest cybersecurity teams may themselves be machines.
FAQ
What is VEXAIoT?
VEXAIoT is a two-agent AI framework—one agent finds vulnerabilities, the other executes attacks—built to test IoT devices in controlled environments like IoTGoat and Metasploitable2.
How successful was the autonomous hacking?
Across 260 attack runs the system succeeded 95% of the time, with 94.5% against IoTGoat and 96.7% against Metasploitable2, most finishing in under two minutes.
Was this tested on real devices?
No. The experiments ran in simulated vulnerable environments, and the authors explicitly frame the result as a security-testing benchmark rather than a deployed attack tool.
