This podcast episode delves into the intricate nexus of artificial intelligence and security, featuring an enlightening conversation with Harriet, the author of a newly released book Practical AI Security. We explore her compelling journey from a background in physics and anthropology to becoming a pivotal figure in the realm of cybersecurity, particularly focusing on the challenges posed by adversarial machine learning. Harriet elucidates the pressing necessity for organizations to comprehend and mitigate the security vulnerabilities inherent in AI systems, as well as the broader implications for national security. Our discourse also addresses the critical need for collaboration between cybersecurity professionals and AI developers to ensure that security considerations are embedded within AI design from the outset. Ultimately, we aim to provide our audience with a profound understanding of the evolving landscape of AI security and the imperative of safeguarding these transformative technologies.
🎙️ Security by Default Podcast
Practical AI Security: Attacking, Defending, and Securing the Future of AI
With Harriet Farlow — Founder of Mileva Security Labs & Author of
Practical AI SecurityArtificial Intelligence is transforming the way we build technology, automate decisions, analyze data, and solve some of the world’s biggest challenges.
But as AI becomes more powerful and more deeply embedded into our lives, one critical question becomes increasingly important:
How do we secure AI itself?
In this episode of Security by Default, host Joseph Carson is joined by Harriet Farlow, AI security researcher, founder of Mileva Security Labs, and author of “Practical AI Security: A Hands-On Guide to Attacking, Defending, and Securing Modern AI Systems.”
Together they explore the rapidly evolving world of AI security, adversarial machine learning, and why understanding how AI works is essential before we can protect it.
About This Episode
AI is often described as the next technological revolution, but securing AI requires us to rethink many traditional cybersecurity approaches.
Unlike conventional software, AI systems are built on data, probability, optimization, and learning models. They do not always fail in predictable ways, and vulnerabilities are not always solved with a simple patch.
Harriet shares her fascinating journey from studying physics and anthropology to working in data science, national security, and artificial intelligence, eventually discovering the world of adversarial machine learning — where attackers attempt to manipulate and disrupt AI systems themselves.
This conversation goes beyond the hype and explores what defenders, developers, and organizations need to understand as AI becomes a critical part of modern technology.
What You Will Learn
🤖 Why AI Security Matters More Than Ever
AI is becoming part of software development, business operations, healthcare, finance, critical infrastructure, and cybersecurity itself.
As adoption accelerates, organizations must move beyond simply asking:
“How can we use AI?”
and start asking:
“How do we secure AI?”
🧠 Understanding How AI Really Works
Harriet explains why machine learning systems are fundamentally different from traditional software.
AI systems are:
- Probabilistic rather than deterministic
- Dependent on training data quality
- Designed around optimization
- Continuously influenced by changing environments
Understanding these foundations is essential for anyone responsible for protecting AI.
🔓 The World of Adversarial Machine Learning
What happens when attackers stop targeting only applications and infrastructure…
…and start targeting the AI model itself?
The episode explores:
- Model manipulation
- Data poisoning
- AI weaknesses
- Training challenges
- Unexpected behaviors
- The difficulty of understanding model decisions
🛠️ How Do You Patch AI?
One of the biggest questions facing cybersecurity professionals today:
If AI learns something wrong, how do we fix it?
Traditional security follows a familiar process:
Find vulnerability → Apply patch → Reduce risk
AI changes that.
Sometimes protecting AI is not about fixing code.
It is about understanding and correcting behavior.
⚔️ AI for Security vs Security for AI
For years, organizations have focused on using AI to improve cybersecurity.
But now the challenge has expanded.
Cybersecurity needs AI.
But AI also needs cybersecurity.
As AI becomes part of everyday systems, security teams must understand how to protect the models, data, and decisions that organizations rely on.
🌍 Why AI Security Requires Different Skills
The future of AI security requires collaboration between:
- Cybersecurity professionals
- AI engineers
- Data scientists
- Researchers
- Risk leaders
- Policy experts
Building trustworthy AI means bringing these worlds together.
Security must be part of AI from the beginning.
Key Topics Discussed
🔹 Harriet’s journey from physics and anthropology into AI security
🔹 Working in data science and national security environments
🔹 Discovering adversarial machine learning
🔹 Founding Mileva Security Labs
🔹 Writing Practical AI Security with No Starch Press
🔹 Why AI vulnerabilities are different from software vulnerabilities
🔹 The importance of data quality and model training
🔹 Understanding probability and machine learning foundations
🔹 How attackers target AI systems
🔹 Why securing AI requires a new mindset
🔹 The future of AI safety and cybersecurity
🔹 Staying updated in a fast-moving industry
🔹 Building responsible and secure AI systems
Memorable Quotes
💬 “Before we can secure AI, we first need to understand how it works.”
💬 “AI security is not always about fixing a bug. Sometimes it is about correcting a behavior.”
💬 “Cybersecurity needs AI, but AI also needs cybersecurity.”
💬 “The future is not just about building smarter AI — it is about building safer AI.”
Episode Chapters
00:00 – Introduction to Security by Default
01:03 – Harriet Farlow’s origin story
04:28 – From data science to cybersecurity
08:48 – Creating Mileva Security Labs
10:51 – Conferences, community, and writing Practical AI Security
17:28 – How AI has evolved
19:43 – Understanding machine learning models
21:43 – The challenge of patching AI systems
23:37 – Training data, quality, and user impact
25:23 – Why AI models can be difficult to understand
27:36 – AI and cybersecurity coming together
30:18 – Why AI fundamentals matter
32:04 – Practical examples and real-world AI security
33:38 – Staying updated in AI security
36:27 – Learning from the AI security community
38:08 – Ethics and responsible AI development
Guest
Harriet Farlow
Founder — Malevra Security Labs
Author — Practical AI Security
🔗 LinkedIn:
https://www.linkedin.com/in/harriet-farlow-654963b7/
📘 Practical AI Security — No Starch Press
https://nostarch.com
🎓 AI Fundamentals Course
https://harriethacks.com/course/
Listen & Subscribe
🎧 Security by Default Podcast
Exploring the people, stories, and ideas helping make technology safer.
Because security should not be an afterthought.
Security should be by default.
#SecurityByDefault #AISecurity #Cybersecurity #ArtificialIntelligence #MachineLearning #AdversarialML #AI #ResponsibleAI #SecurityResearch
Takeaways:
- The podcast episode discusses the importance of understanding AI security in the context of national security and its implications.
- Harriet's journey from a background in physics and anthropology to her current role in AI security demonstrates the interdisciplinary nature of the field.
- The conversation highlights the necessity for collaboration between AI developers and cybersecurity professionals to ensure secure AI systems.
- Listeners are encouraged to engage with various resources to stay informed about the rapidly evolving landscape of AI and cybersecurity.
- The significance of addressing the ethical considerations in AI development is emphasized throughout the discussion, focusing on empowering rather than replacing human effort.
- The episode underscores the idea that AI security is not merely about using AI for cybersecurity but also about securing AI systems from external threats.
