====== Learning to Hack AI ======

When I started this blog, my goal was simple: teach myself how to become a hacker and document the journey as I go.

At first I kept things broad. I explored Linux, networking, programming, and general cybersecurity concepts. That phase was important because hacking isn’t really about one specific skill. It’s about understanding how systems work well enough to see where they break.

But lately I’ve been thinking about where I want to focus my attention.

The answer keeps pulling me in the same direction: **artificial intelligence**.

AI systems, especially large language models, are being integrated into everything. Search engines, customer support, coding tools, research assistants, and automation platforms are all starting to rely on them. Whenever technology spreads that quickly, it creates a new landscape for security research.

And that landscape is still very new.

===== Why AI Security Interests Me =====

Traditional hacking focuses on things like servers, networks, and software vulnerabilities. AI introduces an entirely different kind of system to analyze.

Instead of just code, you're dealing with:

  * statistical models
  * massive training datasets
  * probabilistic outputs
  * instruction systems and guardrails

That means AI systems have **new kinds of weaknesses** that didn’t exist before.

Understanding those weaknesses is becoming a new frontier in cybersecurity.

===== What I Want to Learn =====

My current goal is to understand how modern AI systems work at a deeper level. Specifically, I want to focus on the mechanics behind large language models.

Some of the things I’m studying include:

==== Tokens and Tokenization ====

Large language models don't see text the way humans do. They break language into smaller pieces called tokens. Understanding tokenization helps explain why AI sometimes behaves strangely and why certain prompts work better than others.

==== Context Windows ====

AI models can only "remember" a certain amount of information in a conversation. This limit affects how instructions, prompts, and conversations influence the model's behavior.

==== Guardrails and Safety Systems ====

Most AI systems include guardrails that try to prevent misuse or harmful outputs. These guardrails are implemented through a combination of training methods, filtering systems, and prompt engineering.

Understanding how those guardrails work is an important part of understanding the system itself.

==== Adversarial Prompting ====

Researchers have discovered that AI systems can sometimes be influenced in unexpected ways through carefully constructed prompts. Studying these behaviors helps reveal how AI interprets instructions and how its internal reasoning processes work.

==== Using AI as a Tool ====

AI isn't just something to analyze. It's also a powerful tool. Large language models can help with research, coding, automation, and analyzing complex information.

Learning how to use AI effectively as part of a technical toolkit is becoming an increasingly valuable skill.

===== Why This Matters =====

AI is quickly becoming part of the infrastructure of the internet. As more systems rely on machine learning and language models, understanding how these technologies work will become increasingly important for both security professionals and researchers.

Just like web applications needed security researchers in the early days of the internet, AI systems will need people who understand how they behave under pressure.

That’s where my curiosity is taking me right now.

===== What I Will Explore =====

Going forward, this wiki will likely cover topics like:

  * how large language models process language
  * the concept of tokens and context windows
  * how AI guardrails are designed
  * examples of adversarial prompting in research
  * using AI tools for learning and technical research

I’m still at the beginning of this path. There’s a lot to learn and a lot to experiment with.

But that’s what hacking has always been about.

Curiosity.  
Experimentation.  
Understanding systems deeply enough to see what others miss.

Right now, the system I want to understand is **AI itself**.