What are the main risks of AI agents?

There are five real workplace risks worth knowing. They sound dramatic because they are, but they are all preventable. Shadow AI is staff using unapproved tools. Data leakage is pasting sensitive data into those tools, where one in five organisations have reported a breach this way. Hallucination laundering is signing your name to confident but false AI output. Prompt injection is attackers hiding malicious instructions in documents the AI later reads. Rogue agents are autonomous tools left running after a project ends. A customer support desk that adopts an unapproved tool to speed up replies, then starts feeding it live tickets containing customer data, is shadow AI and data leakage happening at once, with no record of what left the building.

Are AI agents safe to use at work?

They can be, but only with real governance in place. Using AI without governance isn't innovation, it's risk, and banning it outright just drives it underground. Safety comes from knowing four things: which tools your team is actually allowed to use, what data each one can reach, who is responsible if something goes wrong, and how you keep an eye on what's actually running. An HR team wanting to screen applications with an AI tool is using it safely when that tool has been formally evaluated for security and compliance first, not just allowed onto the network because it happened to work. Start by finding out which tools in your team have been formally evaluated, not just what's technically allowed on the network.

What is prompt injection?

Prompt injection is when an attacker hides malicious instructions inside a document or email that your organisation's AI later retrieves and processes. Those hidden instructions can override the safeguards you built into the tool, which makes it a serious risk. Imagine an ops team running an AI assistant that reads incoming supplier emails to draft replies: a booby-trapped email could carry instructions the assistant follows without anyone realising, steering it past the limits you set. Any AI that reads outside content can be steered by that content, so review what your deployed tools are allowed to access and act on.

How do you put guardrails on an AI agent?

Guardrails are really about governance rather than a single setting. In practice that means deciding four things: which tools are allowed in the first place, what data each one can reach, who is answerable if something goes wrong, and how you keep track of what is actually running, so a proof-of-concept agent doesn't keep operating as a forgotten backdoor. Picture an agent spun up to help with finance month-end, then left running once the project closed: without monitoring, it stays connected to your systems long after anyone is watching it.

Using agents safely: risks and guardrails

// in this post

What this means for you
Try this
Common questions about AI agent risks

// actions

↗ share on linkedin

The main risks of AI agents and how to put guardrails on them: shadow AI, data leaks, hallucination laundering, prompt injection, and rogue agents.

// next in the series

09The future of AI agents (and what to do now)13 min watch

// ai agents

Get the next lessons as they drop

New lessons land in batches. Subscribe and I'll email you when the next one goes live.