VEXAIoT: Autonomous IoT Vulnerability EXploitation using AI Agents

Technical Breakdown§

Core Methodology§

VEXAIoT leverages two LLM-based agents—a Vulnerability Detection Agent (VDA) and an Attack Execution Agent (AEA)—that collaborate via a shared state and a tool library (Nmap, Metasploit, searchsploit, etc.). The VDA performs reconnaissance and identifies potential vulnerabilities (e.g., OWASP IoT Top 10). The AEA then plans and executes exploits using a chain-of-thought reasoning loop. Each agent's LLM (GPT-4 or similar) is prompted with a system prompt that defines available tools and a structured output format (e.g., JSON with action and parameters).

Architecture§

class VEXAIoT:
    def __init__(self, llm):
        self.vda = VulnerabilityDetectionAgent(llm)
        self.aea = AttackExecutionAgent(llm)
        self.state = {"scan_results": None, "vulnerabilities": []}
    
    def run(self, target_ip):
        self.state = self.vda.analyze(target_ip)
        for vuln in self.state["vulnerabilities"]:
            attack_plan = self.aea.plan(vuln, self.state["scan_results"])
            result = self.aea.execute(attack_plan)
            # Update state based on result
        return self.state

Key Implementation Details§

• Tool Integration: Each tool is wrapped as a function with a natural language description. The LLM outputs JSON specifying the tool and arguments.
• Memory: A persistent conversation history with the LLM stores previous actions and results, enabling context-aware decisions.
• Evaluation: 10 attack scenarios (e.g., default credentials, command injection, buffer overflow) in IoTGoat (a deliberately vulnerable IoT firmware) and Metasploitable2. Each scenario run 26 times.

Mathematical Formulation§

The attack success rate is defined as: $$ \text{Success Rate} = \frac{\# \text{successful exploits}}{\# \text{all attempts}} \times 100\% $$ VEXAIoT achieves $95.0\%$ overall (260 attempts), with $94.5\%$ on IoTGoat and $96.7\%$ on Metasploitable2.

Token overhead is measured as average tokens per attack, reported as "low" (under 4K tokens per scenario). Average execution time is under 2 minutes for most attacks, including LLM inference and tool execution.

Results Summary§

• OWASP IoT vulnerabilities covered: weak passwords, insecure network services, insufficient authentication, etc.
• Failure cases: mostly due to LLM hallucination (wrong tool arguments) or environment instability (e.g., service not responsive).