Skip to content
0xbenzo
Back to projects

AgentScan AI

Technical Overview

AgentScan AI is a static analysis tool designed specifically for testing LLM applications. Unlike classic code parsers, AgentScan parses tool schemas, input sanitation code blocks, and prompts to compute a vulnerability score.

Key Capabilities

  • Abstract Syntax Tree (AST) Inspection: Recursively checks Python files to locate agent definitions and highlight integrations where system calls are initiated.
  • Simulated Injection: Generates adversarial payloads using local weights to test tool definitions for instruction override safety.
  • Secure Sandbox Running: Safely simulates tool calls within isolated environments, recording prompt leak traces.