Practical guide
How to secure AI agents before they reach production.
An AI agent becomes more capable when it can use tools, access business data, and complete tasks across applications. Those capabilities also introduce decisions that need careful thought.
A useful starting point is to define what the agent is allowed to access, which actions it can perform, and where additional checks are needed.
Define what the agent is supposed to do.
Before deciding which controls to implement, define the agent’s task and boundaries.
For example, an agent that summarises support tickets may need to read ticket content. It may not need permission to delete tickets, issue refunds, or change customer accounts.
A clearly defined purpose makes unnecessary permissions and risky actions easier to identify.
Questions to answer
- What task is the agent responsible for?
- Which data sources does it need?
- Which tools can it use?
- Which actions can it perform?
- Which actions require approval?
Ticket summary agent
Example scope
Needs
- Read ticket contentRead only
Doesn’t need
- Delete ticketsNot granted
- Issue refundsNot granted
- Change customer accountsNot granted
Five decisions that make a difference.
- 01
Give it only the access it needs.
Use the smallest set of permissions that allows the agent to complete its task.
Example
A reporting agent may need read access to a specific dataset, not permission to change the underlying records. - 02
Protect the information it handles.
Decide which data the agent can retrieve, process, retain, and send to other systems.
Example
An assistant preparing a sales summary should not automatically receive unrelated employee or customer records. - 03
Control how it uses tools.
Validate tool requests and enforce authorisation in the application rather than relying on the model’s response alone.
Example
An agent may prepare an email, while a separate check controls whether it can actually send it. - 04
Add approval where it matters.
Use human review for sensitive actions or operations that are difficult to reverse.
Example
An agent can prepare a change request, but an authorised person confirms the final change. - 05
Test and monitor behaviour.
Test unexpected inputs and review activity to identify unusual behaviour, failures, and opportunities to improve the controls.
Example
Test whether the agent follows malicious instructions embedded in an external document.
From request to approved action.
Imagine an AI agent that prepares a customer account update. The model can help interpret the request, but the surrounding application should still validate the requested change and confirm the user’s authority.
Select a step to read its explanation.
Illustrative example. The appropriate checks depend on the system and the action being performed.
The system receives the user’s task or instruction.
The agent identifies the action it believes is needed.
The application verifies that the requested operation is permitted.
Sensitive actions are sent to an authorised person for review.
The application performs only the permitted operation.
The outcome and relevant activity are recorded for appropriate review.
New tools mean new questions.
An agent’s risk profile can change when it receives new permissions, connects to additional data, or gains the ability to take external actions.
Review its controls when you change its model, tools, prompts, integrations, or responsibilities.
Ask yourself
- Can it access information it does not need?
- Can it perform actions without sufficient checks?
- Are external instructions handled as untrusted input?
- Are sensitive actions subject to appropriate approval?
- Can unusual activity be detected and investigated?
Need help planning an AI workflow?
SingleBit builds software products and AI-enabled applications. Tell us what your agent needs to do, which systems it connects to, and where you need help with the engineering.
We can discuss your requirements and explore possible next steps within the team’s confirmed capabilities and the scope of the engagement.
Tell us about your AI project.
Share a few details about your agent, its integrations, or the workflow you want to improve. Let’s discuss your requirements in a free 30-minute consultation.