Security /

AI

Think past the model

The model is one piece. The data, tools, connectors, memory, and identities around it matter just as much.
The useful security boundary is the whole system, not just the prompt box.

Agents

First question: what can the agent reach?
Second: what can it actually do with that access?
Anything that can take an action needs tighter permissions than something that only returns text.

MCP

MCP is another integration boundary.
Check the server, the tools and data it exposes, which identity is being used, and what that identity can do.

Prompt injection

Content is input, not authority.
Reading a webpage, file, or message should not automatically grant permission to act on instructions inside it.

Red flags

Data and tool access.
Approval points before sensitive actions.
Secrets, logging, third-party access, and where permissions cross systems.