During her tenure in the Australian intelligence community, Harriet Farlow observed a notable trend: discussions regarding artificial intelligence (AI) and cybersecurity predominantly concentrated on enhancing security measures through AI, while considerably fewer conversations revolved around the imperative need to secure AI systems themselves. This disparity frustrated Farlow, especially considering that many organizations within Australia heavily rely on AI models developed outside the country. Concerned about the security vulnerabilities that such dependencies could introduce, she set out to address this critical oversight in her recently published book, Practical AI Security.
In Practical AI Security, available through No Starch Press, Farlow, who currently serves as the CEO of AI security consultancy Mileva Security Labs, offers a comprehensive technical foundation for securing AI systems. The book covers essential topics, including the fundamentals of machine learning models, navigating governance frameworks, and implementing red-teaming campaigns to proactively identify potential weaknesses.
In an enlightening interview with TechTarget, Farlow elaborated on the myriad challenges associated with AI system security. She emphasized the necessity for Chief Information Security Officers (CISOs) to establish robust guardrails and define a clear level of acceptable risk when it comes to AI deployment. Unlike traditional security paradigms, AI security introduces a series of unique challenges that warrant special consideration.
Farlow delineated several key factors that differentiate AI security from conventional security measures. First is the probabilistic behavior of AI systems; whereas deterministic systems yield consistent outputs from the same inputs, AI often produces different results despite identical input conditions. This inherent uncertainty complicates risk assessment and management. Next, the heavy reliance on data raises critical questions about the completeness and quality of the data used to train AI systems, along with concerns about potential biases and backdoors that could be hidden within the training datasets. Lastly, as AI models become increasingly capable, especially in the realm of agentic AI, they may inadvertently function in a manner that poses insider threats to organizations, demanding vigilant oversight.
"These elements introduce a slew of security quandaries that differ from those faced by traditional deterministic systems," Farlow remarked, highlighting the complexities encountered in securing AI environments.
Despite these multifaceted challenges, Farlow is adamant that organizations should not shy away from leveraging AI technologies. Instead, the focus should shift toward identifying conditions that permit the safe utilization of AI. She urged CISOs to reframe their perspectives, advocating for an inward-looking approach that prioritizes securing AI systems from within, rather than only guarding against external threats.
One significant concern Farlow discussed is the phenomenon of Shadow AI, a situation where employees resort to using unauthorized AI tools, such as personal accounts for applications like ChatGPT, when organizational policies are overly stringent. "Striking the right balance between security and usability is always a challenge," she noted. "If your security measures are too restrictive, staff will explore alternatives to get their work done."
Even in scenarios where employees adhere strictly to security protocols, the probabilistic nature of AI makes it improbable to create an infallible security framework. "The reality is that there’s always some degree of risk you will need to accept. It’s essential to define what that risk threshold looks like," Farlow explained.
When it comes to establishing AI guardrails, Farlow emphasized that there is no universal approach. Organizations may choose to appoint a Chief AI Officer with exclusive decision-making authority or adopt a decentralized structure that distributes responsibility across various roles. The critical factor is whether those entrusted with oversight possess the visibility required to monitor all interaction points of an AI system, as well as the authority to halt potentially risky AI initiatives.
Maintaining open communication among the various stakeholders involved in AI deployment—ranging from security teams and IT staff to AI developers and third-party vendors—is another fundamental aspect of robust AI governance. CISOs must actively assess the security risks associated with external AI tools, aware that organizations often exert less control over these resources.
Farlow pointed out that meaningful governance should not be an impediment to innovation. Instead, it serves as a framework that enables organizations to proceed at a sustainable pace while still mitigating risks. "Throughout my experience, I’ve witnessed organizations eagerly adopt new technologies, despite potential security implications," she recounted. "It’s perplexing when I see companies rushing into AI deployments without considering what those systems can access. It’s a recurring theme."
As the landscape of cybersecurity evolves, Harriet Farlow’s insights underscore the importance of comprehensive AI security strategies that account for the unique challenges presented by AI technologies. The ongoing dialogue within the industry will surely benefit from her expertise and commitment to fostering safer AI environments, as illustrated in her work with Mileva Security Labs, alongside her various platforms that promote intelligence around AI security.
Katherine Finnell, the senior editor for TechTarget Cybersecurity, has played a critical role in publishing essays that delve into complex topics within the field, including the intersection of AI and cybersecurity strategies.

