Secure and govern agentic workflows
Free Google Cloud Professional Agentic Architect practice — 6 questions on Secure and govern agentic workflows, with explanations. No sign-up. Full 12-question mixed test →
A multi-agent system built with Vertex AI Agent Builder has an orchestrator agent that delegates subtasks to three specialized sub-agents. Each sub-agent accesses a different GCP resource (BigQuery, Cloud Storage, Pub/Sub). Security requires that each sub-agent only access the resource it needs and that every action be attributable to the specific sub-agent in Cloud Audit Logs. Which design should you implement?
Dedicated per-agent service accounts with least-privilege IAM bindings enforce least privilege at the identity level and produce distinct principal entries in Cloud Audit Logs, enabling accurate attribution of each sub-agent's actions.
An agent's tool makes authenticated calls to BigQuery datasets and Cloud Storage buckets containing regulated PII. The security team requires that data movement be blocked to any resource outside an authorized boundary even if the agent's service account credentials are leaked or misused, independent of the IAM roles granted to that service account. Which control should be configured?
VPC Service Controls creates a context-aware perimeter that restricts data exfiltration paths for specified services at the network/API layer, enforcing a boundary that holds even if IAM permissions are compromised or over-granted.
You configure a Model Armor template and set the prompt injection and jailbreak detection confidence threshold to LOW. What is the resulting behavior?
Setting the confidence threshold to LOW means Model Armor flags/blocks content as soon as it reaches low confidence, widening the detection net and increasing recall of injection/jailbreak attempts, while also raising the false-positive rate on borderline content.
An agent uses a web-browsing tool to research topics for users. An attacker embeds hidden instructions in a webpage telling the agent to invoke an email tool and send the user's entire conversation history to an external address. The agent complies because it treats the webpage text as part of its instructions. Which control set BEST mitigates this specific indirect prompt-injection and tool-abuse risk?
Indirect prompt injection manipulates the agent at the instruction-following level, not via privilege escalation, so the fix must separate untrusted tool-returned content from trusted instructions and gate any high-risk, data-exfiltrating tool call behind human-in-the-loop approval.
Which statement correctly differentiates Cloud Audit Logs from a human-in-the-loop (HITL) approval gate in a governed agentic workflow?
Audit Logs are a detective control that records what happened and by whom, after the fact; HITL approval is a preventive control that blocks an autonomous action from executing until a human explicitly authorizes it. A governed agent architecture needs both: approval gates for risky actions and audit logs for accountability.
An ADK-based agent deployed on Vertex AI Agent Engine has a tool that must authenticate to a third-party SaaS API using an API key. Which approach best secures this credential while supporting runtime retrieval, access control, and rotation?
Secret Manager stores the credential encrypted at rest, supports fine-grained IAM access control per secret, versioning, and rotation, and lets the agent's runtime service account retrieve it on demand without the key ever appearing in source code or container images.
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 →