Start with the task
Identify what the system can read, change and send. The risks of drafting a paragraph differ from the risks of changing a customer record or controlling equipment.
Protect the boundaries
Keep secrets out of prompts and browser code. Separate trusted instructions from retrieved content. Give tools only the permissions they need.
Keep changes recoverable
Use previews, approvals and logs for consequential actions. Make it possible to reverse mistakes and stop a workflow that is behaving unexpectedly.
Key takeaways
Safety is built from several layers: good instructions, narrow permissions, careful handling of data, useful evaluation and human oversight.