Ask the Experts: Gen AI, Cybersecurity, & AI Agent Questions Answered

Differentiating Generative AI and Agentic AI

A common point of confusion is the distinction between generative and agentic AI. While related, their core functions differ significantly.

• Generative AI is fundamentally about creation. As its name implies, it generates new content—text, code, images, or music—in response to a specific prompt. It operates in a reactive mode, waiting for user input and then producing an output. Key use cases include copywriting, image generation, and summarizing existing content.

• Agentic AI is focused on achieving a goal. Instead of a simple prompt-response interaction, an agent is given a high-level objective. It then autonomously plans, decides, and executes a series of multi-step actions to reach that goal without continuous human intervention. This process involves several stages:

  1. Planning: Devising a strategy to meet the objective.
  2. Execution: Calling upon large language models (LLMs) or other domain-specific tools to perform tasks.
  3. Memory: Maintaining context and remembering the results of previous actions to inform future steps.
  4. Feedback Loop: Continuously self-improving and adapting its approach based on new information and changing context.

Use cases for agentic AI are more complex and process-oriented, such as creating autonomous incident response runbooks in cybersecurity or advanced robotic process automation (RPA) that can adapt on the fly.

Understanding AI Hallucinations

When an AI model confidently states false information as fact, it’s referred to as a “hallucination.” This isn’t an intentional act of lying, as there is no intent, but rather a byproduct of how LLMs function.

LLMs are prediction machines, not knowledge databases. Their primary function is to predict the most statistically likely next token (roughly, a word) in a sequence. They are optimized for fluency and cohesion, not factual accuracy. This can lead them to fill in knowledge gaps with plausible-sounding but incorrect text.

Several factors can increase the likelihood of hallucinations:

• Recent Events: Queries about topics that post-date the model’s training data cutoff.
• Niche Topics: Subjects where there is very little training data available for the model to draw from.
• Leading Questions: Prompts that suggest an answer, which the model may then adopt and elaborate on.

While larger, more advanced models have reduced the frequency of hallucinations, the risk is inherent to their architecture. The primary mitigation strategy today is Retrieval-Augmented Generation (RAG), where a model is fed relevant, factual information from an external source (like a vector database) at the time of the query to ground its response in reality. However, for now, a “human-in-the-loop” approach for validation remains essential to ensure the accuracy of AI-generated content.

Cybersecurity in the Modern Age

The Dark Web: More Complex Than It Seems

A common question is why the dark web can’t simply be blocked or outlawed. The reality is multifaceted:

1. Difficulty in Finding: The “dark” web is so-named because it is not indexed by search engines. It is estimated to be less than 2% of the total web content, making it incredibly difficult to locate comprehensively.
2. Jurisdictional Challenges: The internet is global. While one country might outlaw certain content, it can simply be moved to a server in a different jurisdiction, making enforcement nearly impossible.
3. Dynamic Nature: The dark web operates like a game of “whack-a-mole.” Sites can appear, disappear, and relocate quickly, making any attempt to block them a constant chase.

Furthermore, there is a debate about whether completely eliminating the dark web would be desirable. It serves as a vital tool for free speech and journalism in oppressive regimes. For cybersecurity professionals, it is an invaluable resource for monitoring the activities of malicious actors to understand their tactics and tools.

The Danger of “Zero-Click” Attacks

It is a misconception that you are safe as long as you don’t click on anything malicious. A class of threats known as zero-click attacks can compromise a system simply by visiting a compromised website.

This is possible due to vulnerabilities in the complex software that makes up a modern web browser and its components:

• Browser Bugs: Browsers are intricate applications, and like all complex software, they have bugs. A certain percentage of these are security-related.
• Active Content: Technologies like JavaScript, plug-ins, and browser extensions run code on your machine. While this code is supposed to be confined within a “sandbox” in the browser, a security vulnerability can allow it to escape and execute malicious commands on the host operating system.

Until software becomes perfect, visiting any website carries a degree of risk. Users should be cautious about the sites they visit, as a “bad neighborhood” online can have serious consequences.

The Future of AI and Employment

The question of whether AI will take jobs is a significant concern. The more likely scenario is that AI will transform jobs rather than replace them entirely, much like ATMs shifted the role of bank tellers from dispensing cash to more complex customer service tasks.

Most jobs are a bundle of various tasks. AI is currently well-suited to automate some of these tasks, but not necessarily an entire role.

AI excels at:

• Repetitive, pattern-based tasks.
• Large-scale data processing.
• Creating “first drafts” of documents or code.

AI currently struggles with:

• True creativity and empathy.
• Complex, nuanced reasoning.
• Physical dexterity and adapting to novel situations.

A job may be more vulnerable to AI automation if it scores highly in three areas:

1. It is highly rule-based and deterministic.
2. It involves extensive documentation or writing that requires little judgment.
3. It consists of low-context tasks that do not require a deep understanding of surrounding, unstated factors.

For professionals whose roles fit this description, now is a good time to focus on upskilling and developing capabilities in areas where humans still hold a distinct advantage.

• Tags:
• Agentic AI
• Generative AI
• Cybersecurity
• RAG
• Hallucination
• Zero-Click Attack
• Dark Web
• AI Careers