DevBox

README excerpt

# AgentSecure - AI Agents Security Scanner (Multi-LLM, OpenSource Frameworks)

AgentSecure is a security review system for AI agents. It analyzes agent prompts, tool bindings, MCP configurations, dependency policies, and project files before those agents are deployed.

The project was designed for the Google I/O hackathon around a hybrid Gemini/Gemma architecture: use fast cloud models for routing and reasoning, local models for sensitive data, managed sandboxes for controlled execution, and human approval before any remediation is applied.

## Problem

AI agents are increasingly defined by system prompts, connected tools, MCP servers, dependencies, browsing permissions, and repository access. That creates a new security review problem:

- Prompt injection can override the agent's intended behavior.
- Tool access can expand an agent's blast radius.
- Agent prompts may contain credentials, PII, or sensitive business logic.
- MCP and dependency configuration can introduce unsafe capabilities.
- Security teams need audit evidence, not just a model-generated opinion.
- Single-provider scanners create blind spots because one model is judging another agent's behavior alone.

AgentSecure treats AI agent security as an adversarial, multi-model review workflow rather than a static prompt linting task.

## Solution

AgentSecure is a hybrid AI agent security composer. It routes each part of an agent review to the best available model or scanner, runs attacks in a controlled sandbox, gen