Breaking Down What Secure By Design Means For Generative AI

 

Just this past Thursday, I gave a presentation to about 150 people (coworkers of mine) about Generative AI and Cybersecurity.  I touched upon a few use cases, and the kinds of questions that a client or a prospect might ask if they are interested in this kind of solution. 

At the end of the presentation, a question was asked to me:  “How do I safeguard my PII (Personal Identifiable Information) data if I know I am submitting to an application that uses Generative AI?”

I had to think about this one, because obviously I did not know the answer.  There is no silver bullet answer to this one, so this can be difficult to answer.  As far as I know at the present time, there are no concrete controls in place for this. 

But rather, the security must be baked in as soon as you start the process of building your Generative AI model and eventually deploying into your production environment. 

This concept is technically known as “Secure by Design”, and it can be defined as follows:

“Secure by design is a philosophy and approach that prioritizes security considerations at every stage of the development lifecycle, ensuring that systems and products are inherently secure from the start. Instead of adding security as an afterthought, secure by design integrates security measures throughout the entire design process, making it a foundational element of the product or system.”

(SOURCE:  Google Search)

In other words, security is not after though after deployment of anything, but rather, it is planned out from the very beginning to make sure that nothing has been left behind.  This has started to become the mantra in the world of software development, but now, given the evolution of Generative AI and its long-term potential, it is now being deployed here as well.

This is a concept that has been strongly advocated by CISA, and to read a comprehensive white paper, click on the link below:

http://cyberresources.solutions/Blogs/CSIA_SBD.pdf

This is an initiative-taking approach to help secure your Generative AI models, right from the very beginning.  Because of the confusion of what means “secure” in this realm of the world, at the present time, there is a reactive approach that is being taken to this.  Here are some of the consequences:

1)     Data Poisoning:

One of the cardinal rules in Generative AI is that you want all the datasets that you will be using to be as “cleansed” as possible, meaning that they are free from any or other erroneous bits of information.  But in this regard, the Cyberattacker, and unknowing to you, can easily insert some kind of malicious payload into them.  This will eventually lead to data exfiltration or data leakages happening.

2)     Prompt Engineering:

This is the art of Generative AI in which it teaches you how to create specific queries to get the best answers (or outputs) that are possible.  It takes time to learn all of this.  But, the Cyberattacker is stealthy enough that they can inject malicious prompts into your Generative AI model, thus having it give out answers that it should not be given. 

3)     Deserialization:

This is the process where the Generative AI model can take bits of data and transform them into other types, especially when they are needed for software development.  Once again, a malicious payload can be deployed here by the Cyberattacker to create a back door to get into the software application after it has been deployed.

So how can Secure by Design come to work here?  It does it primarily by adopting the principles of what is know as “Machine Learning Security Operations”, or “ML” SecOps for short.  This is where the Generative AI, the Operations team, and the IT Security team come together in one unison to make sure that the Generative AI model has security baked into right from the very beginning of the development lifecycle.

Some of its components are as follows:

Ø  Threat Modeling:  This is where future threat variants are predicted, to help beef up the lines of defenses.

 

Ø  Data Preparation:  This is where all efforts are made to ensure that all the collected data sets are cleansed and optimized, as stated earlier.

 

Ø  Testing:  You want to evaluate the Generative AI model as it is being developed throughout its various stages.  A tool that can be used here are what are known as “model scanners”.

 

Ø  Continuous Monitoring:  Even after the Generative AI model has been deployed into the production environment, you will still want to keep an eye on it and monitor any rogue network traffic that may come its way.

 

Ø  Penetration Testing:  This is where you will want the Red and Blue Teams to launch comprehensive exercises so that any gaps or vulnerabilities can be found and quickly remediated.  You do not want to wait until the end of this.

 

Ø  Firewalls:  One of the best tools that you can use to protect your Generative AI model is using something like a Next Generation Firewall.  These are much more sophisticated versions of the traditional Firewall and can analyze data packets at a much more granular level.

 

Ø  Incident Response:  In this regard, you will want to have playbooks that can be automatically triggered to help contain a security breach to your Generative AI model should it ever occur.

My Thoughts on This:

Even despite using the concepts of MLSecOps, you must have bought it from every key stakeholder that participates in the development of every aspect of the Generative AI development process.  This event also includes C-Suite. 

They must know what is going on, and they cannot plea ignorance or that they have not been informed.  There must be a Change Management Committee that will oversee any changes to the Generative Model after it has been developed and deployed into the production environment.

So, in the end I guess, the best answer for tight as to how to safeguard your PII is just as initiative-taking on your end as the MLSecOps team would be as well.  Trust your gut.  If something does not feel right, do not enter your private and confidential information into it.