Restricting an AI Assistant's Tool Permissions is a practical security course for learners comfortable entering terminal commands. Protect a document assistant while preserving its useful reading workflow.
Two Guided Labs teach operation allowlists and document ownership/sharing. One Challenge combines these skills in a new project. Proposals are openly labeled deterministic replays; authentication, document reads, and server authorization are real. No live model is used.
What You Will Learn
- Limit a role to the tools its task needs, including direct handler requests.
- Enforce owner and shared access using the authenticated principal.
- Reject forged identity arguments while preserving legitimate reads.
- Edit small JSON and Python policies and interpret allowed/denied responses.
Who This Course Is For
For beginning security and application developers with basic terminal familiarity. JSON, Python policy syntax, and authorization concepts are taught before use. Each unit provides an independent, prepared Linux environment with Policy Editor, Test Client, and Assistant Tool Lab interfaces.
Frequently Asked Questions
Do I need an AI account or GPU? No. Replayed proposals require neither.
Does this prove prompt-injection resistance? No. It practices deterministic server-side tool permissions, not live-model behavior.
Will I write code? Yes: small explained JSON and Python edits. The Challenge uses the same mechanisms and supplies a separate complete solution. Assessment checks actual permissions and preserved functionality, without reports or copied logs.
