AI Agent Security with kosa8: Sandboxes, Undo, Governance

About this course

This course contains the use of artificial intelligence.

AI coding agents are useful, and they go wrong in predictable ways: a destructive command, a prompt injection that leaks data, an untrusted tool, an exposed secret, or no record of what happened. This course teaches you to run agents so that those failures are contained, reversible and on the record.

You will learn the threat model first: what an agent is, how indirect prompt injection works, why ordinary containers share the host kernel, and why the controls that matter are the ones an agent cannot talk its way around. Then you will put it into practice with kosa8, a local-first container platform that runs each sandbox as its own microVM.You will create sandboxes with network egress enforced on the host, mount only the folder an agent needs, take checkpoints that capture memory and running processes, and rewind in seconds when something breaks. You will fork a live environment into parallel copies to try several fixes at once. You will run Claude Code, Codex, Gemini CLI, Copilot CLI and other agents from ready-made, least-privilege sandbox kits, and write your own.

Next you will connect agents to kosa8 over the Model Context Protocol, run third-party MCP servers in their own sandboxes behind a gateway, serve local models on OpenAI- and Anthropic-compatible endpoints, and drive remote workers.

For teams, you will sign and enforce an organisation policy, approve MCP tools per team, require named approvers with a two-person rule, label and encrypt sensitive environments, seal snapshots, forward a hash-chained audit log to a SIEM, and export evidence packages a third party can verify offline. The final section tours sector packs (GxP validation, malware analysis, cyber ranges, disconnected edge sites, CUI labelling and AI assurance) and ends with a capstone project.

Read the full description on Udemy on Udemy (opens in a new tab)

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.