An agent that can call tools can also be talked into calling the wrong ones. Prompt injection hides instructions in a web page, email or tool result; a model can leak personal data in an answer; generated code can touch files it should not. Security for agents means limiting what each agent can reach and checking what goes in and out, so one bad input cannot do much damage.
These guides cover the layers in order. Guardrails check inputs and outputs for injection, personal data and policy breaks. A dedicated guide goes deeper on prompt-injection defence. Then come sandboxes for running model-written code, verifiable identity for each agent, isolation between customers in a multi-tenant product, and an audit log that records who did what. The MCP security guide applies the same ideas to tool servers.
The code in these guides runs on Promptise Foundry, an open-source Python framework for AI agents and MCP servers.