Governance, Safety, and Risk Management
Free Claude Certified Architect - Professional practice — 6 questions on Governance, Safety, and Risk Management, with explanations. No sign-up. Full 12-question mixed test →
A financial services firm is building an agent that uses MCP servers to query customer accounts and execute trades. They must minimize the risk that a prompt injected into a tool's response (e.g., malicious text embedded in a market-data feed) causes an unauthorized trade. Which design BEST mitigates this risk while keeping the agent functional?
Human-in-the-loop confirmation for irreversible, state-changing actions creates a system-level control boundary that a prompt injection embedded in tool output cannot bypass, while still allowing autonomous read-only operations — this is the core least-privilege/defense-in-depth pattern tested for MCP tool design in regulated workflows.
A healthcare provider plans to send patient PHI to Claude via the Anthropic API to summarize clinical notes. Which step is REQUIRED before this is compliant with HIPAA?
HIPAA requires a covered entity to have an executed Business Associate Agreement with any vendor that processes PHI on its behalf; this is a contractual/legal prerequisite, not a model configuration, and must be in place before any PHI is sent to the API.
A German insurance company requires that none of the model input/output data leaves EU-based infrastructure at any point, per internal data residency policy. Which deployment approach BEST satisfies this requirement?
Data residency is an infrastructure/region guarantee — deploying via a hyperscaler's EU region ensures inference and storage stay within that geographic boundary, which is how EU data residency requirements are actually satisfied for Claude deployments.
An architect is designing a customer-support agent that must never reveal its system prompt contents or internal tool names even under adversarial user prompting. Which configuration choice provides genuine defense-in-depth at the system/application level, beyond the model's own training?
Defense-in-depth at the system level requires controls that operate outside the model's own reasoning — an output filter that catches leaked secrets/identifiers and sanitized error handling form an independent layer that still protects even if an adversarial prompt bypasses the model's instruction-following.
Which statement best differentiates Anthropic's Responsible Scaling Policy (RSP) / AI Safety Level (ASL) framework from a system-level guardrail that an architect would implement in their own application?
Anthropic's Responsible Scaling Policy defines AI Safety Levels (ASL) as organization-level commitments about safety and security practices required before training/deploying models of a given capability, which is distinct from — and complementary to — the deployment-time controls an architect implements in their own application.
A regulated bank must demonstrate to auditors a complete, tamper-evident record of every prompt and completion made by their Claude-based underwriting assistant, including which internal employee triggered each request. Which approach BEST satisfies this auditability requirement?
Regulatory audit requirements for identity attribution and tamper-evidence can only be met by a customer-controlled, append-only audit logging system built at the application layer, since the vendor's own abuse-monitoring logs are not designed for, or guaranteed to retain, per-employee attribution or long-term compliance evidence.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →