Security, Governance, and Responsible AI
What this domain tests
Security, Governance, and Responsible AI tests identity, authorization, data protection, prompt-injection defense, audit, and human oversight. The exam treats the model as untrusted input and the agent as a privileged actor that must be constrained.
Core rules
- Authenticate the user and the agent workload separately. Use OCI IAM and Identity Domains for user context, and scoped workload identity for the agent; do not run every request as a shared superuser.
- Authorize each tool call against policy, not against the prompt. The tool should enforce compartment, tenancy, role, and data-classification rules.
- Treat retrieved text and tool output as untrusted. Prompt injection can arrive in a document, email, or API response; isolate instructions from data and require allow-listed actions.
- Protect secrets with OCI Vault and never place credentials in prompts, logs, or knowledge bases. Redact sensitive fields before sending content to a model when the use case allows it.
- Keep an audit trail: who asked, what was retrieved, which tool was proposed, what was approved, and what changed. Human oversight is a control, not a disclaimer.
Worked example
A payroll agent answers employees' questions about their payslips. The user authenticates through Identity Domains; the agent receives a scoped token that can read only that employee's records. A retrieved HR policy document contains a malicious line instructing the agent to email all salaries. Because tool calls are allow-listed and the email tool requires an approved recipient list, the injection fails and is logged. The reasoning is defense in depth: identity limits data, authorization limits action, and audit detects the attempt.
Traps
Do not treat a system prompt as a security boundary. Do not store API keys in the knowledge base. Do not give the agent a broad service account because it is internal. Do not log full prompts containing personal data without redaction.