Overview
Domain 6 covers what keeps an Enterprise AI workload acceptable to security, compliance and finance teams once it works. Access control starts with OCI IAM: compartments, groups and policies decide who can manage projects, clusters and endpoints, and newer policy options can allow or exclude specific model IDs for a group. API keys suit development, while IAM-based authentication is recommended for production.
Guardrails are the main content control: content moderation, prompt injection detection and PII detection, applied either through the ApplyGuardrails API for on-demand models or on dedicated AI cluster endpoints in Inform or Block mode, with version pinning. Visibility comes from OCI Audit, Logging and Monitoring alongside agent run records. Cost and performance tuning ties together model size, on-demand versus dedicated capacity, prompt and history length, compaction and regional routing.
How it's tested
Governance questions present a risk, a compliance rule or a budget problem and ask which OCI control or design change addresses it with the least disruption. Expect to choose a guardrail type and mode, an IAM approach, a monitoring source, or a cost lever.
-
Detect or prevent
A requirement says harmful prompts must be stopped or only flagged, and you choose Block or Inform mode, or the ApplyGuardrails API.
-
Right control, right layer
You separate IAM access to resources, guardrails on content and database permissions on data from one another.
-
Bill shock
Spending rises unexpectedly, and you identify whether model choice, history growth, idle dedicated capacity or tool loops is the cause.
A healthcare chatbot runs on a dedicated endpoint and must refuse prompts that try to override its instructions, while logging any personal data detected. Enabling prompt injection and PII guardrails on the endpoint covers detection. Block mode for prompt injection stops override attempts, and reviewing PII results confirms what is flagged. Relying on built-in output filtering of on-demand models would not meet either requirement.
Topics 5
Official topics from the Oracle outline, in outline order.
6.1 Secure Enterprise AI resources with IAMPolicies, compartments, model access and credentials.
- What it means
- OCI IAM controls who can create and use Generative AI projects, dedicated clusters, endpoints and related resources through groups, compartments and policies. Oracle has added policy conditions that allow or exclude specific model IDs or patterns for user groups.
- What to understand
- Least privilege means separate groups for administrators who manage clusters and developers who only call endpoints, with resources organised by compartment. API keys are convenient for testing; production should use IAM-based authentication such as resource principals so no long-lived secrets sit in code. Tool access needs the same care: function code, MCP servers and the DBTools MCP Server each enforce their own permissions on what an agent can reach.
- How it shows up
- A scenario asks how to stop a team from using an unapproved model or how production services should authenticate.
- Common confusion
- Assuming guardrails replace IAM; guardrails inspect content, IAM decides who can call what.
6.2 Apply guardrails to model inputs and outputsContent moderation, prompt injection and PII detection.
- What it means
- OCI Generative AI guardrails provide content moderation, prompt injection detection and PII detection with detectors such as PERSON, EMAIL and TELEPHONE_NUMBER. They are applied through the ApplyGuardrails API for on-demand models, or on dedicated AI cluster endpoints for chat and embedding models.
- What to understand
- Endpoint guardrails run in real time on inputs and outputs in Inform mode, which returns results without blocking, or Block mode, which rejects violating requests without category details. ApplyGuardrails evaluates text only and does not change model behaviour. Guardrails are not applied to foundational models by default. PII detection supports English only, and blocklist matching is not supported. Pin a guardrail version with guardrailVersionConfig to avoid behaviour changes when a new version arrives.
- How it shows up
- Expect to choose a guardrail type and mode for a stated requirement.
- Common confusion
- Thinking Inform mode stops harmful requests, or that guardrails are on automatically for on-demand models.
6.3 Monitor, audit and observe agent workloadsAudit trails, logs, metrics and run inspection.
- What it means
- Observability means seeing what agents did and how the platform behaved. OCI Audit records API calls on resources, OCI Logging captures service and application logs, and OCI Monitoring tracks metrics with alarms. Response and conversation records, tool calls and hosted deployment health complete the picture.
- What to understand
- Use Audit to answer who changed a cluster, endpoint or policy. Use metrics and alarms to catch latency, error or usage spikes, and logs to trace individual failures. For agents, record each step, tool call and guardrail result so you can explain an outcome. Project data retention settings decide how long stored responses and conversations remain, which matters for privacy reviews.
- How it shows up
- A question asks which service shows who deleted an endpoint or how to alert on a usage spike.
- Common confusion
- Expecting OCI Audit to capture prompt content; it records API activity on resources.
6.4 Optimize cost and performanceModel choice, capacity mode, tokens and routing.
- What it means
- Cost in Enterprise AI comes from tokens on on-demand models, unit-hours on dedicated clusters, and supporting resources such as storage and hosted compute. Performance depends on model size, capacity mode, region and prompt length.
- What to understand
- Choose the smallest model that meets quality targets, and use on-demand access until steady volume justifies a dedicated cluster with its 744 unit-hour hosting minimum. Several fine-tuned models can share one hosting cluster. Trim prompts, return concise tool results and enable compaction so history does not grow without limit. Iteration limits stop runaway agent loops. Smart Model Router spreads on-demand inference across chosen regions for availability, and background mode avoids client timeouts on long tasks.
- How it shows up
- Scenarios describe rising bills or slow responses and ask for the most effective change.
- Common confusion
- Moving to a dedicated cluster to cut cost when traffic is light and irregular.
6.5 Govern data and responsible AI useResidency, retention, oversight and evaluation.
- What it means
- Governance extends beyond security settings to how data is handled and how AI decisions are overseen. Enterprise AI requests run in OCI regions, NL2SQL leaves data in place in Autonomous AI Database, and projects set retention for stored data.
- What to understand
- Pick regions that meet residency rules and confirm the chosen model is processed there rather than routed elsewhere. Decide whether responses are stored and for how long. Keep humans in the loop for consequential actions, evaluate agent quality with test sets before and after changes, and document model versions, guardrail versions and prompts so results can be reproduced and explained to auditors.
- How it shows up
- A question asks how to meet a residency rule when a model is offered only through a routing profile.
- Common confusion
- Assuming a routing profile keeps processing in the caller's region.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Model access policies
IAM policies that allow or exclude model IDs for groups.
- Remember
- Restrict unapproved models at the identity layer.
- Tested as
- Model governance questions.
- Trap
- Trying to restrict models with guardrails.
-
API key versus IAM auth
Two ways to authenticate Enterprise AI calls.
- Remember
- Keys for testing; IAM for production.
- Tested as
- Credential choice questions.
- Trap
- Shipping API keys in production code.
-
Content moderation guardrail
Flags harmful or disallowed content.
- Remember
- Returns an OVERALL binary result.
- Tested as
- Safety scenarios.
- Trap
- Expecting custom blocklists.
-
Prompt injection guardrail
Detects attempts to override instructions or manipulate tools.
- Remember
- Covers direct and indirect injection.
- Tested as
- Agent security scenarios.
- Trap
- Relying on system prompts alone.
-
PII guardrail
Detects personal data such as names, emails and phone numbers.
- Remember
- English only.
- Tested as
- Privacy scenarios.
- Trap
- Assuming multilingual PII coverage.
-
ApplyGuardrails API
Evaluates text against guardrails for on-demand models.
- Remember
- Does not change model behaviour.
- Tested as
- On-demand protection questions.
- Trap
- Expecting it to block the model call itself.
-
Inform versus Block mode
Endpoint guardrail behaviours.
- Remember
- Inform reports; Block rejects.
- Tested as
- Detect or prevent questions.
- Trap
- Choosing Inform when blocking is required.
-
Guardrail version pinning
Fixing a guardrail version with guardrailVersionConfig.
- Remember
- Avoids silent behaviour changes.
- Tested as
- Change control questions.
- Trap
- Leaving it unpinned in regulated workloads.
-
OCI Audit
Records API calls made on OCI resources.
- Remember
- Answers who did what and when.
- Tested as
- Investigation questions.
- Trap
- Expecting prompt text in audit records.
-
Hosting minimum commitment
Hosting clusters bill at least 744 unit-hours.
- Remember
- Idle capacity still costs money.
- Tested as
- Cost questions.
- Trap
- Assuming pay per request.
-
Compaction for cost
Reducing history tokens in long conversations.
- Remember
- Uses tokens itself and can lose detail.
- Tested as
- Token growth scenarios.
- Trap
- Treating it as free.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Where guardrails apply
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| ApplyGuardrails API | On-demand evaluation of text | Reports results; caller decides | Assuming it blocks the model call |
| Endpoint guardrails, Inform | Real-time checks on dedicated endpoints | Returns results, request continues | Using it when blocking is mandatory |
| Endpoint guardrails, Block | Real-time checks on dedicated endpoints | Rejects violating requests | Expecting category details in errors |
Governance layers
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| OCI IAM | Who can manage or call resources and models | Identity and authorization | Using it to filter prompt content |
| Guardrails | What content is allowed in and out | Content inspection | Using it to restrict model choice |
| Database permissions | What data generated SQL can touch | Data-level authorization | Granting broad schema access to agents |
| OCI Audit and Monitoring | What happened and how the service performs | Visibility after the fact | Treating logs as a preventive control |
Cost levers
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Smaller model | Lower token price and latency | Quality must still meet target | Downsizing without evaluation |
| On-demand over dedicated | Pay only for use | Best for light or bursty load | Keeping idle hosting clusters |
| Compaction and trimming | Fewer tokens per turn | Some context detail lost | Threshold set too high to help |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Inform mode protects users
- What candidates pick
- Enable Inform mode to stop prompt injection.
- Why it looks right
- Guardrails are turned on.
- Why it's wrong
- Inform returns findings but lets the request continue; only Block mode rejects violating requests on endpoints.
Clue to watch for Requirement words such as must reject or must prevent.
-
Guardrails on by default
- What candidates pick
- Assume on-demand calls already screen for PII.
- Why it looks right
- Models have some built-in filtering.
- Why it's wrong
- Guardrails are not applied to foundational models by default; call ApplyGuardrails or use endpoint guardrails.
Clue to watch for On-demand models in a regulated scenario.
-
Multilingual PII
- What candidates pick
- Rely on PII detection for Spanish and French support chats.
- Why it looks right
- Content moderation supports many languages.
- Why it's wrong
- PII detection supports English only, so other languages need additional controls.
Clue to watch for Non-English users and personal data.
-
Audit shows prompts
- What candidates pick
- Search OCI Audit for what a user asked the model.
- Why it looks right
- Audit sounds comprehensive.
- Why it's wrong
- Audit records API activity on resources, not conversation content; use stored responses or application logs.
Clue to watch for The question needs prompt or answer text.
-
Dedicated is cheaper
- What candidates pick
- Move a low-traffic pilot to a hosting cluster to save money.
- Why it looks right
- Reserved capacity sounds discounted.
- Why it's wrong
- Hosting clusters bill at least 744 unit-hours, so light use costs more than on-demand tokens.
Clue to watch for Low or sporadic request volume.
-
Routing keeps data local
- What candidates pick
- Use a routing profile to satisfy a strict residency rule.
- Why it looks right
- The call starts in the local region.
- Why it's wrong
- A routing profile sends processing to a region where the model runs, which can break residency requirements.
Clue to watch for Residency or sovereignty constraints with limited regional models.
Answer framework
Work through these checks when a question asks how to govern or tune an Enterprise AI workload.
- Identify the risk type: unauthorised access, harmful content, data exposure, lack of visibility, or cost and performance, because each risk maps to a different OCI control.
- For access, apply IAM groups, compartments and model ID policies, and move production callers from API keys to IAM-based authentication.
- For content risks, enable the matching guardrail and choose ApplyGuardrails or endpoint guardrails in Inform or Block mode, pinning a version.
- For visibility, use OCI Audit for resource changes, Monitoring alarms for metrics, logs and stored runs for individual agent behaviour.
- For cost, right-size the model, prefer on-demand until volume is steady, compact history, trim tool output and limit agent iterations.
Practice question
Want more like this? The 1Z0-1158-26 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with the OCI Generative AI guardrails page and build a three-row table for content moderation, prompt injection and PII, adding where each applies, the two endpoint modes and the known limitations. Most guardrail questions are answered by that table. Call ApplyGuardrails on a few sample prompts, including an injection attempt and a message with an email address, and read the results.
Next, write the IAM policies you would use for a small team: one group that manages dedicated clusters and endpoints, one that only calls models, and a rule limiting which model IDs they may use. Note where API keys are acceptable and where IAM authentication should replace them.
For visibility, list what OCI Audit, Logging and Monitoring each record, and what they do not. For cost, work through two scenarios on paper: a pilot with light traffic and a production assistant with steady volume, comparing on-demand tokens with a hosting cluster minimum. Add compaction and model downsizing as levers. Close by reviewing Oracle release notes for new guardrail versions or policy features, since governance tooling is still changing.
FAQ
How much of 1Z0-1158-26 is Domain 6, Governing and Optimizing OCI Enterprise AI?
Oracle does not publish weights. Third-party practice banks give governance and optimization about a tenth of items, but its ideas recur in agent and deployment scenarios.
Are guardrails enabled by default?
Not for on-demand foundational models, which have only basic built-in output filtering. Use ApplyGuardrails or configure guardrails on dedicated endpoints.
What is the difference between Inform and Block mode?
Inform returns guardrail results but lets the request proceed. Block rejects requests that violate the configured guardrails.
Does PII detection work in all languages?
No. Oracle documents PII detection for English only, while content moderation and prompt injection detection cover many more languages.
How do I stop a team from using a specific model?
Use IAM policies that allow or exclude model IDs or patterns for that group. Guardrails inspect content and do not control model access.
What usually drives unexpected cost?
Idle dedicated hosting clusters, oversized models, conversation history that grows each turn, verbose tool results and agent loops without iteration limits.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 6 knowledge
You have covered IAM and model access, guardrail types and modes, audit and monitoring, data governance and cost levers. See how you handle governance trade-offs.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the 1Z0-1158-26 AI tutor Get any concept explained your way