Page 269 - Cyber Defense eMagazine Annual RSA Edition for 2024

AI Security Challenges

There’s already no shortage of commentary speculating on AI’s concomitant dangers, most of which fall
into one of three broad categories:

1. Training Data-Related

AI systems have voracious appetites for data. Their need to ingest and analyze large and diverse
datasets necessarily means that AI developers run the standard gamut of risks that attach whenever a
sufficiently rich, centralized data asset is created. There are questions of storage, usage, and access.
Data must be protected from external bad actors, of course, but it must also be safeguarded against
accidental unauthorized access or misuse. AI further complicates questions of basic security because AI
agents are increasingly able to make their own decisions about what to do with the data assets to which
they have access.

2. Deep Analysis & Inference

AI technologies have a genuinely unprecedented capacity to analyze massive datasets and make
complex inferences, which can make them into powerful surveillance tools, even inadvertently. We’ve
already seen obvious instances of AI surveillance with facial recognition programs, but the true power –
and risk – of sophisticated AI lies in its ability to draw accurate conclusions about individuals based on a
broad array of seemingly unrelated data points. Such capabilities aren’t simply a risk due to potential bad
actors; instead, these inferential capabilities can inadvertently uncover sensitive information in the regular
course of once-benign activities, like behavioral customer segmentation.

3. Misuse of Capabilities

AI will accelerate virtually every industry, including fraud. We’re seeing it already with cloned voices and
deepfakes – AI can be used to create a distorted view of the world, then distribute that view in an
incredibly broad or highly targeted fashion depending upon the desires of individual bad actors. The
potential attack vectors are too many to count, ranging as they do from email fraud to driving recruitment
for extremist groups. The key takeaway here is that AI can serve as a force multiplier for the most
malignant forms of human creativity.

Beyond these, there’s a long tail of potential threat vectors that could arguably justify the creation of
additional categories, but that discussion is starting to feel a bit academic. For our purposes, it’s
principally important to understand that AI-related privacy and security risks take multiple forms,
recapitulate older vulnerabilities while creating new ones, and can arise from benign as well as malign
intentions. The sheer breadth of possibility here demands a complex, active, and collaborative response.

De-Risking AI: The Story So Far

Just as AI is in its relative infancy as a technology, so too are efforts to comprehensively safeguard
privacy in an AI-driven world. At the regulatory level, we’re seeing efforts from a number of national and
international bodies to begin enshrining privacy protections and somewhat regulate AI.

269