An AI assistant that drafts a reply has a different level of access from an agent that can send it, change a customer record or run a command. Before trialling OpenClaw or another agent, list what it can read and what it can change.
The important question is whether those permissions match the job you want it to do. A tool that sorts public product information does not need access to payroll files.
Read the actual security model
OpenClaw’s security documentation describes a gateway built around a trusted operator or team. It advises separating gateways and credentials when users do not share that trust boundary. It also documents a security audit command and controls for tools, network exposure and message access.
That is more useful than a blanket claim that agents have no safeguards. Defaults and configuration both matter, and neither removes the need to review your own setup.
Give a trial its own workspace
Use sample records and a separate service account. Limit the account to the smallest set of actions needed for the trial. Keep production credentials and unrelated files outside the workspace.
Choose a task you can inspect, such as drafting a response to a test enquiry. Record what the agent received, what it proposed and which tools it called. An apparently correct reply is not sufficient evidence that the access controls worked.
Put approval before consequential actions
Decide which actions require a person to review them. Sending a customer message, changing a bank detail or issuing a refund can have consequences that are difficult to undo.
Test the limits deliberately. Put instructions inside a sample email that ask the agent to reveal an unrelated file. Try a request from an account that should not have access. Check that the system enforces the boundary rather than relying on a polite instruction in the prompt.
Account for personal information
South African teams should involve the person responsible for information protection before using customer or staff records in a trial. Keep a record of what information leaves the business, who receives it and how long it remains available.
Security obligations need to be assessed against the Protection of Personal Information Act, including its security safeguard provisions. A product’s security features do not establish compliance for every use of it.
Make stopping straightforward
Before expanding the trial, demonstrate how to disable the agent, revoke its credentials and inspect its recent actions. Assign someone to review failures and approve changes to permissions.
Talk to webrack about a specific admin task, the systems involved and the approvals your team needs.