Choosing an AI agent

An AI agent is software that carries out tasks for the people who use it. It might read a mailbox, update a record or post a message. Choosing one starts with what it is allowed to do. Public guidance from UK government, the National Cyber Security Centre and others helps with the questions to ask.

What it may do on its own

The Open Worldwide Application Security Project (OWASP) runs a Gen AI Security Project, which names the risk of an agent doing more than it should "Excessive Agency". It describes damaging actions taken in response to unexpected, ambiguous or manipulated output from the model.

OWASP's advice includes limiting each tool and permission to the minimum needed. It also suggests requiring "a human to approve high-impact actions before they are taken". So a buyer can ask which actions the agent can take without a person agreeing, and whether that list can be changed.

Questions for the supplier

The UK government's Guidelines for AI procurement were published on 8 June 2020 and were written for buyers in government. Their ten considerations include "Avoid Black Box algorithms and vendor lock in" and "Consider the lifecycle management of the AI system".

The National Cyber Security Centre (NCSC) publishes "a series of 12 principles, designed to help you establish effective control and oversight of your supply chain". A buyer can use them to shape the questions for an agent's supplier.

What it records

The NCSC's guidelines for secure AI system development were published on 27 November 2023. Their last stage covers a system once it is deployed, "including logging and monitoring, update management and information sharing".

The government's AI Cyber Security Code of Practice calls an organisation that deploys an AI system a System Operator. Its provision 12.1 says "System Operators shall log system and user actions to support security compliance, incident investigations, and vulnerability remediation." Can the buyer see what the agent did, step by step, and keep that record?

Where your data goes

An NCSC blog post from 14 March 2023 asks the questions a buyer can put to a supplier. One is about what you type into the tool: "Is it available to the vendor's researchers or partners? If so, in what form?" The Information Commissioner's Office (ICO) publishes guidance on AI and personal data for businesses in the public, private and third sectors.

Test it on your own tasks first

A buyer can try an agent on a few of its own tasks before signing. The UK AI Security Institute built Inspect with Meridian Labs. It is "An open-source framework for large language model evaluations", and its makers say it can measure agentic tasks among other things. Its guide shows what a structured test looks like.

A security baseline

The AI Cyber Security Code of Practice was published on 31 January 2025. It is voluntary, and it sets out "baseline cyber security principles to help secure AI systems and the organisations which develop and deploy them". A buyer can ask a supplier which of the Code's 13 principles the product follows.