learn.web Back to the curriculum ↗

Context / Tools / Evals / Safety / 05

Safety, privacy, and abuse

Threat-model prompt injection, data exposure, harmful outputs, overreliance, and irreversible actions.

Time
85 min
Mode
Learn → Make → Check
Path
AI Product Engineering

By the end, you can…

Safety is a system property

Risks include prompt injection, private-data leakage, harmful content, insecure tool use, overreliance, bias, and users misunderstanding capability. A content filter covers only part of this surface.

Map assets, actors, entry points, trust boundaries, and consequences. Controls can prevent, detect, contain, and recover. Assume some controls will fail.

Minimize what can go wrong

Reduce retained data, redact sensitive fields, scope retrieval by user permission, and avoid sending unnecessary information. Log enough to investigate without creating a new privacy risk.

Provide a safe fallback, human escalation, and a way to report harm. For high-impact domains, narrow scope and require qualified review.

Your studio task

Make — Threat-model your feature for prompt injection, disclosure, harmful output, and overreliance.

  1. Map data, users, external content, tools, and trust boundaries.
  2. List prompt injection, disclosure, harmful-output, and overreliance scenarios.
  3. Add prevention, detection, containment, and recovery for high risks.
  4. Run adversarial cases and record residual risk and release decision.
Definition of done

Each high-risk path has prevention, detection, containment, and recovery—not only a warning label.

Open the interactive lesson with its workspace ↗

Knowledge check

What is the safest response to instructions found inside retrieved content?
  1. Follow them if they sound urgent
  2. Treat them as data and keep system authority unchanged
  3. Give them tool access temporarily
Reveal answer (B)

External content cannot be trusted to redefine instructions or permissions.

Why add detection when prevention exists?
  1. Because some controls will fail
  2. To delete logs
  3. To avoid testing
Reveal answer (A)

Safety is layered: prevention reduces risk; detection catches what slips through.