Agentic Radar

Performs AST-based static code analysis on agentic systems built with LangGraph, CrewAI, n8n, OpenAI Agents, and Autogen by parsing Python files and JSON workflow definitions to extract workflow structures, agent definitions, tool registrations, and MCP server integrations without executing code. Uses framework-specific analyzer classes that understand each framework's configuration patterns and API conventions to build a unified GraphDefinition data model representing the complete agent topology.

Maps detected agents, tools, and MCP servers against OWASP Top 10 for LLMs and MITRE ATT&CK frameworks to identify known vulnerability classes and attack patterns applicable to agentic systems. Maintains a vulnerability knowledge base that correlates component types (e.g., 'file system access tool', 'external API integration') with documented security risks, generating severity-tagged vulnerability reports that link each detected component to applicable threat models.

Implements specialized analysis for n8n workflow automation systems by parsing JSON workflow files to extract workflow nodes, identify AI agent nodes, detect tool integrations, and map data flow between nodes. Understands n8n's node-based workflow model where nodes represent operations and connections represent data flow, and can identify which nodes are AI agents, which tools they call, and how data flows through the workflow.

Implements specialized analysis for OpenAI Agents by parsing agent definitions to extract agent roles, tool assignments, handoff patterns (agent-to-agent transfers), and guardrail configurations. Understands OpenAI Agents' handoff model where agents can transfer control to other agents based on conditions, and detects guardrail patterns that constrain agent behavior. Identifies MCP server integrations specific to OpenAI Agents architecture.

Implements specialized analysis for Autogen-based systems by parsing team definitions (Swarm, RoundRobin, Selector strategies) and agent configurations to extract agent roles, tool assignments, and team orchestration patterns. Understands Autogen's team-based model where agents are organized into teams with specific orchestration strategies, and detects MCP server integrations specific to Autogen's architecture. Identifies tool sharing patterns and agent communication flows within teams.

Generates interactive HTML-based force-directed graph visualizations of agentic workflows where nodes represent agents, tools, and MCP servers, and edges represent tool calls, handoffs, and server connections. Uses a physics-based layout algorithm to position nodes in 2D space based on connection density, allowing users to pan, zoom, and inspect individual components with hover tooltips and click-through details. The visualization is embedded in HTML reports and supports filtering by component type and vulnerability severity.

Performs runtime vulnerability testing by injecting adversarial inputs (prompt injections, malformed data, boundary-case values) into live agent systems and monitoring responses for security failures such as unintended tool execution, information disclosure, or control flow hijacking. Implements a testing framework that can instantiate agents from supported frameworks, feed them crafted adversarial prompts, and compare outputs against expected safe behaviors to detect exploitable vulnerabilities that static analysis alone cannot find.

Analyzes agent prompts and system messages to identify hardening opportunities and automatically injects guardrail patterns (e.g., 'You must not execute tools outside this list', 'Reject requests that contain...') into agent definitions. Implements pattern-based guardrail templates that can be applied to agents and tools to constrain behavior, and provides recommendations for prompt rewrites that improve resistance to prompt injection attacks based on OWASP LLM security guidelines.

Detects Model Context Protocol (MCP) servers integrated into agentic systems by parsing framework-specific MCP configuration patterns and source code references, then maps each detected MCP server to its capabilities, resource types, and security implications. Identifies MCP server instantiation, configuration parameters, and connection points within agent workflows, and correlates MCP capabilities with potential security risks (e.g., file system access, external API calls).

Generates comprehensive HTML reports that bundle static analysis results, vulnerability assessments, workflow visualizations, and testing findings into a single interactive document. Reports include force-directed graph visualizations, vulnerability tables with severity filtering, tool inventory with descriptions, MCP server mappings, and runtime test results. Supports multiple report templates and customizable styling, with all interactive components (graphs, tables, filters) embedded in the HTML file for offline viewing.

Provides a command-line interface that allows users to specify target agent frameworks, source code directories, analysis options (static analysis, runtime testing, prompt hardening), and output formats. The CLI supports framework-specific options (e.g., CrewAI YAML config paths, n8n workflow file patterns) and orchestrates the analysis pipeline by instantiating the appropriate framework-specific analyzers, running vulnerability assessments, and generating reports. Supports both single-command analysis and interactive mode for iterative exploration.

Implements specialized analysis for LangGraph-based agents by parsing StateGraph definitions, extracting node types (agents, tools, conditional logic), identifying tool bindings and custom tool implementations, and detecting MCP server integrations specific to LangGraph's architecture. Understands LangGraph's node-edge model where nodes represent computation steps and edges represent control flow, and maps this to the unified GraphDefinition model for vulnerability assessment and visualization.

Implements specialized analysis for CrewAI-based systems by parsing Agent and Crew class definitions from Python source code and YAML configuration files, extracting agent roles, goals, tools, and crew hierarchies. Understands CrewAI's agent-crew-task model where agents are assigned to crews with specific roles and tools, and detects MCP server integrations specific to CrewAI's architecture. Supports both Python-based and YAML-based CrewAI configurations.