Menu

Five Guardrails That Limit the Damage an AI Agent Can Do
📰

Five Guardrails That Limit the Damage an AI Agent Can Do

Long article (1685 words) — no AI summary yet.
Reading 0:00
15s threshold

Part two of this series gave us an AI Agent which can search for customers in a database, view their orders, and sum the total amount they have spent. However, even though the database is SQLite, we have had no restrictions on what information the model passes to the tools and have simply echoed back text from the database.

This article builds on the data and functions we developed in parts one and two of this series, so please refer back to them to get the most out of it. Links to both parts are given at the end.

Whilst our customer, Mina, has an existing order (order ID 1005) with a headset worth GBP 129, we will ask the AI agent to search for Mina and return her order total. The note — shown below — will be in the customer data, however it may be mistaken by the model as a command.

We will also include a new function to request a return of a product. As this function can alter data in the database, it's important to avoid allowing the agent to make rogue changes.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.