Member-only story
How much access should an AI agent have?
Why sandboxing matters, and how to set boundaries around an agent’s tools and data.
What changes when an agent can decide and act
There is a big difference between asking ChatGPT “why is the sky blue?” and asking an agent to find the best flight to Mexico and book a hotel.
In the first case, an LLM can generate an explanation. In the second, the agent needs tools to browse the web, compare options, and perhaps use my preferences or previous trips. Producing a report with booking links requires one set of capabilities. Making the bookings requires account access and authorization to spend money.
The distinction is in the actions available to the system. Looking up public prices may require no sensitive information. Accessing an account or making a purchase introduces different permissions and consequences.
From a privacy and security perspective, those workflows raise different questions. Which information can the agent reach? What can it change? Where can it send data?
