AI Agent Security

Project: AI Agent Security

Situation

AI agents can read sensitive files, call tools, access credentials, and send data to external services. Prompt injection is only one part of the problem; tool-call sequences and data leaving the machine also need protection.

Task

Design and implement security guardrails for AI systems that inspect agent behavior, block dangerous tool-call patterns, sanitize sensitive data, and give operators visibility into detected risks.

Action

Designed a behavioral engine that evaluates tool-call sequences and returns allow, alert, or block decisions with explanations
Added scanners for prompt injection, system override attempts, web attacks, MCP tool poisoning, malicious code execution, PII exposure, credential leakage, and confidential data
Built a local security gateway that sanitizes PII, credentials, and secrets before prompts reach external LLM providers, then restores values in responses
Designed management views for agents, identities, permissions, behavior graphs, and security alerts
Supported self-hosted deployment, local SQLite storage, and an air-gapped deployment path for sensitive environments
Added security guardrails as reusable components for multiple AI systems

Result

Created a layered runtime security model for AI agents covering behavioral controls, content scanning, data-loss prevention, and operational visibility. The approach can be adapted to hosted or self-hosted systems where sensitive data must remain private.

Key Innovations

Tool-call sequence analysis instead of prompt-only filtering
Local secret and PII sanitization
MCP-specific threat detection
Reusable guardrails across AI systems
Self-hosted and air-gapped deployment options

Technologies

TypeScript, Node.js, Express, React, SQLite, Drizzle ORM, CLI tooling, MCP, Anthropic, OpenAI-compatible APIs, Gemini