So, you have some AI tools or are thinking about deploying them and want to know a bit about securing them.
You are not alone, but there are significant challenges due to the rapidly growing capabilities of AI, and the issues around new types of vulnerabilities we may not be used to thinking of. This is a very challenging area to attempt to secure, but I hope to point you in the right direction and set you up with some resources.
The first thing to understand is that agents should be treated as Non-Human Identities (NHI). This is a bit obvious because we employ AI to do many of the same tasks that our workforce does, however, they do them faster and oftentimes they are working 24 hours a day, seven days a week. Unlike a service account or some other NHI, these are expected to interact with many different platforms in ways similar to how a human would.
Imagine an agent that could book your travel and knows your seat preference for an aircraft and what hotels you prefer. That sounds great, but under the hood, we are going to have to give this tool the ability to search the Internet for schedules, interact with the various carriers, check your calendar for potential conflicts and likely even make the purchase. That means they have access to several different platforms and possibly even your credit card information. There are so many ways this can go wrong, and you might not even know it until it is too late.
So, when we think about securing agents and AI tools, there is some guidance out there. Check out the following:
- NIST AI Risk Management Framework (AI RMF) – Provides a lifecycle approach to managing AI risk and “incorporating trustworthiness” into systems
- OWASP LLM Top 10 – Identifies critical vulnerabilities like prompt injection and data leakage
- OWASP AI Security Verification Standard (AISVS) – A structured checklist for verifying AI system security
A zero-trust mindset is also critical given the permissions agents will likely have, so it makes sense that permissions should be kept to the minimum. For example, read-only access to email or calendars unless really needed. Short-lived API keys or authorization tokens can lower the risk but may require a user action on occasion to approve some things before they do it. Splitting tasks and permissions between multiple entities can also be helpful. For example, one agent may have read-only access to email and calendars, while another agent is used specifically to update the calendar with new information. That agent would have rights to access the calendar, but nothing else.
Another concern is prompt injection. This is where a prompt provided to an AI entity is modified or contradicts the expected prompt. Instead of the AI receiving a prompt that tells it to view a person’s calendar, it might tell the AI to delete everything on it, or to enter a fake appointment. This could be the result of an attacker changing the prompt or might even be the product of another AI tool making a mistake when it hands off instructions. Either way, the result is not good.
To counter this threat, a person or an AI specifically trained with a security mindset may be used to review the prompt before it is passed on. There are even AI proxy tools that can look at the prompts and instructions for the AI with an eye for malicious intent. As AI tools continue to evolve, look for tools that have been trained to include security checks before they execute commands.
It is also wise to make sure communication channels and data are secure. Encrypting data is just as important between AI tools and platforms, even those that are hosted on-premise, as it is on the open internet. Micro segmenting of the network can allow you to more easily monitor and control network traffic between the specific tools, making it tougher for bad actors to attack.
Finally, make sure your employees are educated about the potential threats from AI and that your organization has clear and concise policies about how and when AI tools can be used. If employees do not understand what is acceptable, they can easily create dangerous situations unintentionally.
These are just a few suggestions. However, new threats and vulnerabilities with AI are being discovered almost weekly, so it is important to stay up to date with new attack methods and to be cautious with AI tool rollouts.
