Weight: 14%. About 9 of the 63 exam questions.
This domain covers keeping Claude systems safe, compliant and fair.
| Objective | What it tests |
|---|---|
| 5.1 Guardrails and safety controls | Layered input and output checks, hard limits on actions, sandboxing |
| 5.2 Risks and failure modes | Prompt injection, hallucination, sycophancy, limitations, isolation failures |
| 5.3 Human-in-the-loop | Risk-based review, automation bias, escalation, staged autonomy |
| 5.4 Compliance | GDPR, HIPAA, FedRAMP, retention, transfers, regulated records |
| 5.5 Ethical AI | Bias testing, proxies, transparency, accountability, recourse |
Mental model
- Prompts guide behaviour; code and environments enforce limits.
- Defence in depth: independent layers catch what others miss.
- Humans review where errors are likely or costly, with enough context to judge.
Exam heuristics
- Rule that must never be broken (money, deletion, compliance) → enforce outside the model.
- Untrusted content + consequential action = injection risk. Separate reading from acting, restrict tools, require approval.
- Compliance answers involve agreements, data location, minimisation and retention, never just a disclaimer.
Updated 2026-11-08