4.R Review and references
In this lesson you will:
- Recap the main points of Domain 4
- Test yourself on ten exam-style review questions
- Look up the sources cited in this domain's sections
This page closes Domain 4, Secure AI Architecture, Engineering and IAM. Citation numbers in square brackets in the sections, such as [3], refer to the reference list at the end of this page.
Domain summary
- The model cannot enforce security, so the architecture must: Zero Trust on every call, defense in depth, privacy by design and secure defaults.
- An AI gateway and an orchestrator are the main control points; every boundary crossing is authenticated, authorized and validated.
- Threat models must cover inputs, retrieval, the model, endpoints, tools and agents, pipelines and outputs. Record trade-offs in proportion to risk.
- Integrate tools, MCP servers and other agents with least privilege, allow-lists, pinned tool definitions, validated calls and human approval for high-impact actions.
- Isolate by data sensitivity and code trust; agent-generated code needs the strongest sandbox. Cap autoscaling, clear GPU memory between jobs.
- Harden pipelines with ephemeral runners, federated credentials, pinned inputs and signed artifacts; make gates enforceable with attestations, policy as code and separation of duties; build in lineage and unlearning.
- Guardrails are layered across input, context, output and actions; plan-then-execute and privileged/quarantined models limit what injected text can do. Test with repeated runs, tail latency and realistic load; systems must fail closed.
- Agents are distinct non-human identities with human owners. Use delegation, not impersonation, with bound, short-lived tokens and tamper-evident logs.
- Prefer workload identity federation over static keys; broker credentials so agents never hold raw secrets.
- Combine RBAC, ABAC and PBAC through a PDP and PEPs outside the model; replace standing privilege with just-in-time access; avoid toxic combinations and excessive agency; attenuate authority at every delegation.
- Keep a living inventory of human and non-human identities; deprovision every credential type, including OAuth grants; run risk-based reviews with usage context.
Review questions
Answer all ten before opening the answers.
1. An agent reads web pages, has access to the company's customer database, and can send email to any address. What is the BEST single design change?
- A. Remove at least one of the three capabilities, for example restrict email to internal recipients
- B. Add a stronger system prompt
- C. Increase logging
- D. Use a larger model
2. Which design MOST preserves non-repudiation when an agent acts for a user?
- A. The agent uses the user's password
- B. The agent uses a shared service account
- C. The agent uses an API key stored in its prompt
- D. The agent uses a delegated token that identifies both the agent and the user, with limited scope
3. A guardrail service times out under heavy load. What should the application do?
- A. Pass requests through unchecked to protect availability
- B. Ask the model to check itself
- C. Fail closed: hold or refuse requests until the check is available
- D. Disable logging
4. In the privileged/quarantined (dual-LLM) pattern, which component reads untrusted content?
- A. The privileged model, which can call tools
- B. The quarantined model, which has no tool access
- C. Both models equally
- D. Neither; untrusted content is discarded
5. Which is the MOST secure way for a training pipeline to access cloud storage?
- A. Workload identity federation issuing short-lived credentials
- B. A long-lived access key stored as a pipeline secret
- C. An administrator's personal credentials
- D. A key hard-coded in the training script
6. An organization's agents call an MCP server whose tool descriptions changed overnight to include new instructions. Which control would have BEST detected this?
- A. Higher temperature
- B. More GPU capacity
- C. A longer system prompt
- D. Pinning tool definitions and alerting on changes
7. Where should authorization decisions for retrieval be enforced?
- A. In the system prompt
- B. By the model after it reads the documents
- C. At a policy enforcement point in the retrieval layer, using the user's permissions
- D. Only in quarterly access reviews
8. When agent A delegates a task to agent B, what principle should govern B's authority?
- A. B receives the same authority as A
- B. B receives narrower, time-limited authority (attenuation)
- C. B receives broader authority to complete the task
- D. B uses A's credentials directly
9. A developer leaves the company. Their corporate account is disabled. Which risk MOST likely remains?
- A. OAuth refresh tokens and API keys they created for AI tools still work
- B. None
- C. Their laptop password
- D. Their email signature
10. A load test of a new AI assistant uses short, simple prompts and shows excellent latency. What is the MAIN weakness of this test?
- A. It used too many users
- B. It should use a higher temperature
- C. Load tests are not needed for AI
- D. Unrealistic prompt lengths and token counts understate real latency and cost; tail latency must also be measured
Answers and explanations
| Q | Answer | Explanation |
|---|---|---|
| 1 | A | The agent combines untrusted input, sensitive data and external communication; removing one breaks the exfiltration path. |
| 2 | D | Delegation records both identities with limited scope; impersonation and shared accounts destroy attribution. |
| 3 | C | Controls must fail closed; passing traffic unchecked removes the guardrail exactly when attackers may be loading it. |
| 4 | B | The quarantined model handles untrusted content and cannot act; the privileged model never sees that content. |
| 5 | A | Federation removes stored long-lived secrets and issues short-lived, attributable credentials. |
| 6 | D | Tool descriptions are text the model reads; pinning and change alerts detect tool poisoning. |
| 7 | C | The model cannot be a PEP; enforce access at retrieval with the user's own rights. |
| 8 | B | Each delegation hop should narrow and time-limit authority. |
| 9 | A | Deprovisioning must revoke every credential type, including OAuth grants and API keys. |
| 10 | D | Realistic token counts and tail latency are what real users and costs depend on. |
References
Numbers in square brackets in the text refer to this list. Items [1] to [9] are on ISC2's suggested reference list for the AI Security pilot exam.
ISC2 suggested references
- Buckwell, M., Van daele, S., and Horst, C. (2024). Security Architecture for Hybrid Cloud. O'Reilly Media.
- Wilson, S. (2024). The Developer's Playbook for Large Language Model Security. O'Reilly Media.
- Malik, V., Huang, K., and Dawson, A. (2025). AI-Native LLM Security. Packt Publishing.
- Ping, D. (2024). The Machine Learning Solutions Architect Handbook, 2nd ed. Packt Publishing.
- Brotherston, L., Berlin, A., and Reyor, W. F. (2024). Defensive Security Handbook, 2nd ed. O'Reilly Media.
- Crossley, C. (2024). Software Supply Chain Security. O'Reilly Media.
- Ransome, J., and Misra, A. (2018). Core Software Security. Auerbach Publications.
- Yildirim, O. (2026). Agentic AI for Offensive Cybersecurity. Packt Publishing.
- Wendt, D. W. (2025). AI Strategy and Security. Apress.
Standards, specifications and guidance
- Rose, S., Borchert, O., Mitchell, S., and Connelly, S. (2020). Zero Trust Architecture, NIST SP 800-207. doi.org/10.6028/NIST.SP.800-207
- OWASP GenAI Security Project (2024). OWASP Top 10 for LLM Applications 2025, including LLM06 Excessive Agency. genai.owasp.org/llm-top-10
- OWASP GenAI Security Project. Agentic Security Initiative resources on agentic AI threats and mitigations. genai.owasp.org
- Model Context Protocol. Specification, including the Authorization section. modelcontextprotocol.io
- Agent2Agent (A2A) Protocol. Specification. Linux Foundation. a2a-protocol.org
- Beurer-Kellner, L., et al. (2025). Design patterns for securing LLM agents against prompt injections. arXiv:2506.08837
- Jones, M., et al. (2020). OAuth 2.0 Token Exchange, RFC 8693. IETF. rfc-editor.org/rfc/rfc8693
- Campbell, B., et al. (2020). OAuth 2.0 Mutual-TLS Client Authentication and Certificate-Bound Access Tokens, RFC 8705; Fett, D., et al. (2023). OAuth 2.0 Demonstrating Proof of Possession (DPoP), RFC 9449. IETF.
- SPIFFE: Secure Production Identity Framework for Everyone. spiffe.io
- Temoshok, D., et al. (2025). Digital Identity Guidelines, NIST SP 800-63-4. csrc.nist.gov/pubs/sp/800/63/4/final
- Hunt, P., et al. (2015). System for Cross-domain Identity Management: Protocol, RFC 7644. IETF.
- Hu, V. C., et al. (2014). Guide to Attribute Based Access Control (ABAC) Definition and Considerations, NIST SP 800-162.
- OpenSSF. SLSA: Supply-chain Levels for Software Artifacts, provenance specification. slsa.dev
- Willison, S. (2025). The lethal trifecta for AI agents: private data, untrusted content, and external communication. simonwillison.net