AI
Enterprise AI Security: 7 Attacks on Your LLM App, and the Layer That Stops Them
Prompt injection, data leakage, tenant breaches — seven real attacks on enterprise LLM apps, and the Mattrx production code that stops each one.
17 min read5
Long-form tutorials and engineering notes on .NET, React, AI/ML, and SaaS architecture.
Prompt injection, data leakage, tenant breaches — seven real attacks on enterprise LLM apps, and the Mattrx production code that stops each one.
The finale — a year of running MCP in production at Mattrx: what held, what broke, what we'd do differently, and the whole series in one mental model.
The MCP servers you built aren't tied to one model. Here's wiring MCP into OpenAI (client-side + the hosted tool) and agent frameworks — one server, any model.
The protocol was never the enterprise blocker — identity was. Here's rolling MCP across a company: provision once, inherit on login, govern centrally.
An agent that 'feels off' tells you nothing. Here's how to make MCP observable — one trace per agent run, structured logs, and the metrics that matter.
A tool that takes minutes shouldn't freeze your agent. Here's streaming, progress notifications, and the async-job pattern for long-running MCP tools.
A tool result can carry the attack. Here's how to secure MCP against prompt injection, poisoned servers, and tool abuse — and break the lethal trifecta.
A valid token proves who an agent is, not what it may do. Here's scoped, least-privilege authorization for MCP tools — identity plus policy.
An API key is not MCP authentication. Here's how to secure MCP servers as OAuth 2.1 resource servers with Entra ID — tokens, audiences, and identity.
Most MCP tools mirror your REST API and confuse the model. Here's how to design task-shaped tools an agent picks right, calls cleanly, and trusts.
The server exposes tools; the client drives them. Here's a production MCP client — discovery, the agent loop, multi-server routing, and reconnects.
A demo MCP server is a weekend; a production one needs typed tools, real error handling, pagination, and health checks. The Mattrx build, end to end.