Start with one useful job
An AI agent becomes useful when it can do something on behalf of a person or business process. That same ability makes its permissions a design decision. Start by describing one job in plain language: the inputs it needs, the decision it supports, and the actions it can take. A research assistant and an agent that changes customer records should not inherit the same access simply because they use the same model.
Draw the boundaries before choosing controls
List the systems the agent reads, the tools it calls, and the destinations it can write to. Include retrieved documents, email, websites, tool descriptions, and logs. OWASP describes how indirect prompt injection can arrive through external content. Treat that content as data to evaluate; it should not become authority to change the agent’s job or permissions.
Give the workflow an accountable identity
Record who owns the agent and who can change its instructions, tools, and access. Use permissions appropriate to its specific task, with a clear way to remove access. Review the difference between acting as a person, acting through a shared service account, and acting through a dedicated workload identity. The choice affects attribution, access reviews, and incident response.
Separate a recommendation from an action
A model can propose an action; the surrounding system should decide whether the action is allowed. Define the approval points for sensitive changes, external communication, and privileged operations. Check the actual target, scope, and parameters at execution time. A well-written system prompt helps establish intent, but it does not replace authorization in the application.
Test the failure paths
Use a bounded test environment to review malicious retrieved content, misleading tool output, missing information, denied access, and unavailable dependencies. Record expected behavior before running the test. Include a stop mechanism and a recovery path. The useful result is a list of observable controls and remaining gaps, not a claim that a model is immune to manipulation.
Make the release decision reviewable
Bring the workflow owner, system owner, security team, and operations team into the same release decision. Keep a short record of approved capabilities, unresolved limitations, monitoring, and who can pause the workflow. Revisit that record when a tool, model, data source, or permission changes. Start with a narrow capability that can be measured and expand it only when the evidence supports doing so.

