AI security

What it includes

• Control what data goes into the model and where it ends up (provider, logs, retention).
• Least-privilege per agent: it can only touch what its task needs.
• Defence against prompt injection and using AI as an attack vector.
• Output filters: stop the agent leaking sensitive data or acting out of scope.

When you need it

• The agent handles personal or confidential customer data.
• You give AI access to systems where a mistake has real consequences.
• You expose an agent to user input that could try to manipulate it.

How it’s measured

• Agents with verified least-privilege vs. broad access.
• Leak or improper-access incidents (target: zero).
• Prompt-injection test coverage before production.

Common mistakes

• Sending sensitive data to a provider without knowing its retention policy.
• Trusting that the user "won’t try to break it": exposed agents get attacked.
• Treating AI security as something bolted on at the end, not designed in.

FAQ

Does AI security replace normal cybersecurity?

No, it complements it. You still need all your usual security; AI security adds defences for specific threats like prompt injection, context leaks and excessive agent access.

What is prompt injection?

It’s when an input (an email, a document, a web page) contains hidden instructions that try to make the agent do something it shouldn’t. You defend against it by scoping permissions and validating where instructions come from.

Related capabilities