The original title is "Build a Basic AI Agent From Scratch: Security II"

Short summary

Part of a series on building AI agents from scratch, this installment focuses on closing security gaps by moving tool execution into a Docker sandbox, adding prompt-injection defenses, and validating tool inputs against schemas. The author presents a six-section threat model covering prompt injection, tool permissions, I/O validation, resource controls, secret management, and observability. Each control is implemented as a separate auditable module in the accompanying GitHub repository.

• •Six-part threat model for AI agents: prompt injection, tool gating, I/O validation, resource limits, secrets, observability
• •Docker sandboxing isolates agent tool execution to limit blast radius from runaway commands
• •Each security control is a separate module (prompt_safety.py, tool_policy.py, sandbox.py, etc.) for easy auditing

Generated with AI, which can make mistakes.