Security expert Adam (Ads) Dawson reminds us that in the face of unrelenting change and technological advancement, the most powerful weapon can be an insatiable curiosity.

This attribute has defined Ads’ career path. For the past decade, as a self-taught professional and ethical hacker, Ads has excelled in all areas of the security industry, and has developed a strong background in network and infrastructure penetration testing and security. Since joining Cohere, Ads has added large language model (LLM) security to his slate. His work on probing applications for vulnerabilities is vital to helping protect end users from potential breaches.

Ads was also instrumental in the development of the Open Worldwide Application Security Project’s (OWASP) LLM AI Cybersecurity and Governance Checklist (known as the “OWASP Top 10 for LLMs”). Since its release, the checklist has become a foundational resource for corporations and governments seeking to understand LLM security. He’s currently part of the OWASP team focusing on updating the list to reflect generative AI’s evolution. The team plans to deliver an in-depth technical iteration in the next version, and Ads will be authoring several of the current top 10 vulnerabilities.

In this discussion, Ads weighs in on how companies can implement best practices for generative AI security, the vital role of red teaming, the importance of cross-functional collaboration, and why it can be an advantage to feel like the dumbest person in the room.

0:00

/3:57

Ads Dawson on Tackling AI Security Risks

An edited version of our conversation follows:

Describe your career journey and the importance of the OWASP community to your development.

I've always had a passion and drive for what I do. I started as a network apprentice, but my curiosity steered me toward network penetration testing. From there, I made my way into application security. For me, LLM security is an ever-changing landscape.

In the early stages of my career, I was an OWASP child. It's a fantastic framework. I’m passionate about contributing to open source and collaborating with people. For me, it's always been important to give back: I spend a lot of time at local meetups as part of OWASP’s Vancouver and Toronto chapters, talking about security and constantly educating myself. I do love a rabbit hole. And I’ve sought to learn, adapt, and build my own arsenal of tools.

What is red teaming and why is it vital in security?

Red teaming stems from a military concept that involves attacking your own systems to be able to understand how to defend them. When it comes to application stacks and generative AI, red teaming is extremely important in testing the boundaries, looking for weaknesses in systems and processes, and identifying areas with elements of risk. You can then take that information and look to defend the application. If we don't red team to find these areas of risk, it could be a game changer for a company embedding generative AI in their application.

What roles should an organization include on its red team?

Machine learning and generative AI have so many different and unique attack vectors, and some of them overlap with traditional application security. Traditionally, you would have application security engineers integrate security into your product. Companies should really have a diverse red team. It's important to draw on perspectives from different areas—from machine learning engineers to researchers and academics.

What’s an effective starting point for companies looking to tackle LLM security?

A lot of times, I find it can be helpful to take a step back and contextualize security. We should be looking at the fundamentals from the past 30 to 40 years. When embedding generative AI into the application stack, we should understand what additional risk it is introducing. There are a lot of frameworks we can use, such as threat modeling, that can help identify big areas of risk. We can then apply mitigation security in layers and defense in depth. I’m personally a big fan of building your own internal frameworks and taxonomies, derived from existing ones (i.e., threat modeling — stride, secure development — OWASP).

The OWASP security checklist can help you get familiar with the topics before taking a deep dive into specific vulnerabilities. While it is by no means a watertight method, it's definitely a fantastic way to introduce yourself to the biggest vulnerabilities right now.

I don't think there will ever be a framework robust enough, or one that includes concise contextual background for an entire LLM application. Organizations can take the OWASP list and then develop their own framework. To do so, they have to understand their environment stack, build and develop their own framework and tools, and define red teaming methodologies.

How should companies implement best practices for LLM security?

At Cohere, we do a lot of work in adapting and using frameworks, and integrating specific security controls. It starts with the deployment types for LLM applications—from SaaS to on-premises to other model providers. With these factors in mind, we can apply specific mitigations to high-risk areas. For example, rate limiting could be effective in SaaS, but it may not be as relevant to on-premises.

It's not the responsibility of product developers to stay ahead of the emerging threats and industry trends. Product owners drive the product, but it's up to us to secure it. So we really need to be aligned. It requires a joint effort to build a safe environment for our customers. That requires collaboration of cross-functional teams. Involving model developers and application security engineers early in the product development process can ensure that organizations understand what they are building and the security measures they need to implement.

What to learn more? Get in touch with our experts today.

Get Started!

Tackling AI security risks