You Asked About AI: Agents, Hacking & LLMs

An exploration of the evolving AI landscape, covering the paradigm shift in cybersecurity due to AI agents, the practicalities of running local LLMs with tools like Ollama and vLLM, and the emerging stack for agent-to-agent communication.

February 17, 2026 4 min read

You Asked About AI: Agents, Hacking & LLMs — source: https://www.youtube.com/watch?v=ekrOZf_eOJM

The New Frontline: AI in Cybersecurity

The landscape of hacking has fundamentally changed with the introduction of advanced AI. A recent, sophisticated nation-state-level hack utilized Anthropic’s Claude Code, an AI coding agent, to effectively find and attack vulnerabilities. This event highlights a critical perspective, echoing the sentiment of former CISA director Jen Easterly:

We don’t necessarily have a cybersecurity problem. We have a software quality problem.

AI agents are exceptionally proficient at exposing low-quality code and security flaws. This capability leads to what can be described as the “democratization of hacking.” Previously, executing a major cyberattack required a team of elite engineers. Now, powerful and relatively easy-to-use AI tools enable bad actors, even those without deep coding expertise, to deploy autonomous agents. These agents can scan for vulnerabilities, write exploit code, and execute attacks 24/7. The paradigm has shifted from a human-versus-human conflict to an ongoing battle of “my AI defense versus your AI offense.”

Running and Deploying LLMs: From Local Prototyping to Production

It is entirely feasible to run powerful machine learning models locally, moving away from relying solely on high-tech cloud environments.

Local Development with Ollama: For local development and experimentation, Ollama is presented as the gold standard. It functions like “Docker for LLMs,” simplifying the process of downloading and running popular models such as Llama, Mistral, and DeepSeek on a local machine. While you won’t be running 400-billion-parameter models, hardware like an RTX 3090 or 5090 makes it entirely possible to work with 7B, 13B, and even 70B parameter models.
Production-Grade Deployment with vLLM: While Ollama is excellent for prototyping, it is not suitable for production environments. It is not built for scale and queues requests sequentially, meaning it would likely fail under the load of multiple concurrent users. For production, an industrial-grade inference engine like vLLM is the recommended solution. VLLM is designed for high-throughput serving, utilizing techniques like continuous batching to efficiently handle hundreds of concurrent requests. The suggested workflow is to prototype and validate ideas with Ollama locally, then deploy to production using vLLM for robust and scalable serving.

Understanding the Modern AI Agent Communication Stack

A new tech stack is emerging that defines how AI agents interact with tools and each other.

APIs (The Foundation): The traditional method where an application communicates with a server. It’s reliable but rigid, requiring custom manual code for every endpoint and tool integration.
MCP (Model-Context Protocol): Described as the “USB for AI,” MCP is a protocol that standardizes how agents connect to tools. Originally from Anthropic and now open-sourced, it allows an agent to plug into diverse tools like Google Drive, Slack, or Postgres without custom glue code. This simplifies the process of providing agents with access to a wide array of capabilities.
A2A (Agent-to-Agent Communication): This is the highest level of the stack, where AI agents communicate directly with other AI agents. Instead of a single agent interacting with a tool, multiple agents can collaborate, negotiate, and form their own complex workflows. For example, one agent could be tasked with generating code, a second could identify a bug in that code, and a third could implement the fix, all without direct human intervention.

To truly grasp these concepts, the best approach is hands-on experience. Building a simple server with a tool, wrapping it in MCP, and then having a couple of agents communicate via A2A can provide more insight in an afternoon than reading documentation for a week.

AI and the Nature of Creativity

While AI models are proficient at generating creative text, their application in music generation raises deeper questions. The songwriting and melodic choices from some music generation models can be surprisingly excellent, even if the synthesized instrumentation isn’t perfect. This leads to a philosophical debate: are these probabilistic models truly creative, or are they just exceptionally good at remixing existing patterns? Furthermore, it forces us to question whether human creativity is fundamentally different from this advanced form of pattern recognition and recombination.