// topic 05
Governance
Security, permissions, risk, privacy, and keeping meaningful oversight of AI behavior as systems move from a demo to something many people depend on.
Overview
Placeholder — governance as an enabler of adoption (not just a blocker), and the four lenses this topic covers: security, permissions, risk/privacy, oversight.
Security & Permissions
Placeholder — least-privilege tool scopes, prompt injection, the instruction-source boundary pattern.
| Risk | Mitigation |
|---|---|
| Prompt injection via untrusted content | Treat tool/web output as data, not instructions |
| Over-broad tool access | Scope tools per task; require confirmation for irreversible actions |
| Credential exposure | Never place secrets in prompts or logs |
Risk & Privacy
Placeholder — data classification, what should never enter a prompt, vendor/model data handling terms.
Oversight
Placeholder — human-in-the-loop checkpoints, audit trails, escalation paths when AI behavior is wrong.
Open question: where should a human approval gate live so it catches real risk without becoming rubber-stamping?
Running notes
Placeholder — day-by-day notes go here.