First, Do No Harm: Responsibly Applying Artificial Intelligence
- By Nathan Dula
- Dec 10, 2025
It was 2022 when early LLMs (Large Language Models) brought the term “AI” into mainstream public consciousness and since then, we’ve seen security corporations and integrators attempt to develop their solutions and sales pitches around the biggest tech boom of the 21st century. However, not all “artificial intelligence” is equally suitable for security applications, and it’s essential for end users to remain vigilant in understanding how their solutions are utilizing AI.
When a vendor raises the topic of their solution’s new AI capabilities, reactions from potential customers will likely be mixed, depending on their corporate policies, protocols, and personal experiences. Some end users are eager adopters, persuaded by the real benefits they’ve seen from other AI-based technologies, while others are hesitant; their minds filled with memories of hallucinated answers from LLMs. Most are somewhere in-between, aware that AI is bringing real value to many but also nervous about its potential hazards. Few customers are more justified in this nervousness than those in the security industry, where hallucinations and bad data could lead to dangerous complacency, unpredictable security, missed threats – and grave consequences. When trying to understand the potential risks of artificial intelligence for your security ecosystem, it may be helpful to evaluate it across three categories: the developmental, the surface-level, and the open-ended.
Developmental AI is the use of machine learning, deep learning, or other artificial intelligence technologies to build the solution before it reaches the end user. When used for R&D, AI is an incredible tool to speed up and refine painstaking testing and data crunching. This application of AI is considered generally “safe and good” - nearly unanimously well-regarded in the security industry. We say “safe” because it is monitored and vetted by Subject Matter Experts (SMEs) who can confirm the reliability of its outputs, and “good” because it greatly enhances the power of engineering teams to perform rapid iteration on otherwise painstaking tests as well as enhancing the number of scenarios they can use to test their products. While products that use developmental AI can reasonably be called “AI-powered,” they have steady, tested algorithms that are usually locked in to produce the consistent baseline that security customers expect from a product designed for reliable protection.
Surface-level AI, by contrast, operates on the surface of the technology in question. “Surface-level” does not mean unhelpful or lacking sophistication – it just means that its functionality operates away from the fundamental algorithms of the security system that form the baseline for operational capabilities, at the level of user experience. This is the optimal home for the versions of AI that have captured the imagination of the public – LLMs and generative AI. These systems can be useful for reporting and many end-user facing functions, serving as an instant-response help desk to make technical interfaces easier to navigate. Even if this AI can take actions on behalf of the end user, it cannot alter the baseline or create drift in the effectiveness of the technology over time. A safe and good superficial AI in a security product heightens ease of use and accessibility without compromising the fundamentals of detection, deterrence, or data integrity.
Open-ended AI is where a security system can get into trouble. Open-ended AI is when the review and development process of AI deployment extends into the world of the end user, effectively using them for ongoing refinement of baseline performance or data collection. End users, however- unlike the engineers that develop these systems - are not SMEs. They are experts in their own fields, but they are ill-equipped to manage data classification and analysis tasks. So when systems begin inviting them to classify identified threats or nonthreats on the promise of evolving security, the best outcome is that customer inputs are discarded. One dangerous possibility is that the vendor is using the end user’s self-report to alter their detection algorithms, which could cause a decay in system effectiveness over time as the user’s results drift further from the baseline. Another is that the vendor is using their customers’ self-reported data to train future products, which means that poorly vetted data is now threatening the baseline itself.
The key to knowing whether AI in your security systems is good is understanding how it affects the baseline. The baseline can have many adjustable parameters, but the baseline performance is the performance that was used to pass any security certifications that technology has, the performance that was approved by the vendor’s engineers before it left the factory floor, and the performance that all of the vendor’s security promises to you at point of sale are based on. If it changes, you may have evolved your security, but you will lose the ability to trust your security. The ramifications for safety and liability are obvious. When it comes to security, let the innovations of AI make your life easier and your technology better, but don’t let it undermine the consistency you depend on to keep your facility and people safe.