First, let’s go over a few basic AI terms. It’s likely you know many of these, but look over these terms to make sure you understand what we mean by them:
Machine Learning (ML): AI based on learning from data (rather than being specifically programmed to perform a task).
Neural network: An ML method that uses interconnected nodes to learn from patterns and make predictions, using an approach loosely inspired by the human brain.
Deep learning: An ML approach using a neural network with many layers.
Model: A set of data and possibly corresponding programs, trained to identify data patterns to make predictions or decisions. ML systems typically train a model with large amounts of data, then later use that pre-trained model to repeatedly make predictions (“inferences”). Models are sized by the number of parameters; larger sizes tend to be better but require more memory and computation.
Frontier model: A model that is among the currently best available.
Large Language Model (LLM): A deep learning model with many parameters trained to summarize, translate, and generate language. Text inputs and outputs of LLMs are split into tokens (word fragments). See [0xkato2026] for a technical explanation of how LLMs work.
AI chatbot: An AI system that’s designed to interact with a human but cannot perform actions that affect the external environment.
AI agent: An autonomous AI system that can perceive its environment, plan, and execute multi-step actions that affect its environment using external tools to achieve a specific goal.
Agent harness: The program that runs an AI model as an agent, giving it tools and instructions, and running it in a control loop (to perceive, reason, act, and observe). Examples include Claude Code, Goose, and Pi. An agent harness can split work across multiple agents. Don’t confuse this with a fuzzing harness (a small program that feeds generated inputs to the code being tested) or a test harness.
Agent skill: A document (sometimes with supporting resources like programs) that tells an agent how to perform a specific task. Many agent systems load a skill only when it’s relevant to the current task.
AI systems are not sentient. LLMs, for example, repeatedly generate likely next tokens (word fragments); they don’t “understand” in the sense that humans do. Yet scale matters. With many layers and parameters, modern AI systems can simulate intelligence, sometimes astonishingly.
There’s also strong evidence that AI models have improved. One way to measure AI models is the “50%-task-completion time horizon” defined as the “time humans typically take to complete tasks that AI models can complete with a 50% success rate” [Kwa2025]. As of 2025, “this metric has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months” [Kwa2025-blog].
Quiz
Q1. What distinguishes a “frontier model” from other AI models?
It uses symbolic logic instead of a neural network
It can only run on local, organizational hardware
It has no trainable parameters
It is among the best currently-available models
Show answer
Answer: D
Quiz
Q1. What is the key difference between an “AI agent” and an “AI chatbot”?
An agent uses deep learning, while a chatbot uses only hand-written rules
An AI agent can execute multi-step actions using external tools to affect its external environment; a chatbot can’t affect an external environment
An AI agent needs no training data, while a chatbot needs large datasets
A chatbot can perceive its environment, while an AI agent can’t
Finding and Fixing Vulnerabilities Using AI