Pentagon security chief calls for faster AI innovation

When it comes to leveraging the latest advances in AI, the primary challenge is collecting enough data to train a large language model (LLM). A general-purpose LLM such as ChatGPT was trained by pulling data from sources across the Web. When exposed to new data via, for example, retrieval augmentation generation techniques, it then becomes possible for those types of LLMs to, among other similar capabilities, summarize a cybersecurity incident report.

However, with advances in reasoning engines that enable long-term planning, it’s also becoming possible to assign tasks to LLMs that have been specifically trained using a narrow set of vetted data. For example, Cognition Labs is previewing Devin, a generative AI tool that software engineers can assign tasks to complete.

Cybersecurity vendors are naturally researching how to enable similar capabilities to automate cybersecurity tasks, but it’s not apparent how quickly they will be able to deliver on that promise. In theory, providers of cloud security services should be able to aggregate enough data to train an LLM, but as noted by McKeown, those advances are not being made as aggressively as they are in other sectors. The issue, at the moment, is not so much whether LLMs will be used to automate cybersecurity processes but rather when and to what degree.

Arguably, an AI arms race among cybersecurity vendors has been underway for years. Many already make available a wide range of AI capabilities based on predictive and causal AI models that are already widely used. The race to provide generative AI capabilities is now moving beyond providing more advanced probabilistic capabilities enabled by generative AI. The thing to remember, however, is generative AI is designed to surface the right probable answer or recommendation. In an area where being precise matters, generative AI will not likely subsume the need for cybersecurity professionals anytime soon.

Nevertheless, roles within cybersecurity teams will undoubtedly change and evolve in the age of AI. Many of the routine tasks that are often handled by entry-level staff will be eliminated as the overall amount of toil required continues to decline. More experienced cybersecurity professionals will be able to assign tasks to digital assistants that previously might have been performed by a more junior member of the team.

More significantly, organizations that previously would not have been able to mount a serious cybersecurity defense relying on smaller staffs might now be able to. In fact, the chronic cybersecurity skills shortage that has been a major issue for as long as cybersecurity professionals can remember might abate.

As William Gibson famously observed, the future is here; it’s just unevenly distributed. The rate at which it is arriving, however, is faster than most cybersecurity teams today can fully appreciate.