Large Language Model (LLM)
A large language model (LLM) is an artificial intelligence system trained on vast amounts of text to understand and generate human language. LLMs learn by predicting tokens, the smallest units of text, across billions of examples. The result is a model that can answer questions, summarise documents, write code, translate languages, and hold extended conversations with a fluency that earlier AI systems could not approach.
The “large” in the name refers to scale: the number of parameters in the model, the size of the training dataset, and the compute required to build it. GPT-4, Claude, Gemini, and Llama are all LLMs. They differ in architecture choices, training data, and safety measures, but share the same fundamental approach.
In cybersecurity, LLMs are being embedded in development tools, security platforms, customer-facing applications, and internal workflows at a pace that has outrun most organizations’ ability to assess what that means for their risk posture. OWASP GenAI Security Project GenAI Security Project GenAI Security Project published the Top 10 for Large Language Model Applications specifically to address this gap.
How LLMs Work
LLMs are built on the transformer architecture, a neural network design that uses self-attention to process entire input sequences simultaneously rather than one token at a time. Allowing the model to understand context across long passages and generate coherent, contextually appropriate output.
Training happens in two main phases. In pre-training, the model processes enormous text corpora and learns to predict missing or next tokens. This forces it to develop broad representations of language, facts, and reasoning patterns. In fine-tuning, the pre-trained model is adapted for specific tasks or made to follow instructions more reliably, often using human feedback to shape its responses.
The result is a system that is highly capable but also probabilistic. LLMs do not retrieve facts from a database. They generate the most statistically likely continuation of the input they receive. This distinction matters enormously for security, because it means LLM outputs can be fluent, confident, and wrong.
Why LLMs Create New Security Challenges
LLMs do not fit neatly into existing application security frameworks. They accept natural-language input, generate natural-language output, and sit between users and the systems they interact with, creating attack surfaces that traditional controls were not designed to address.
The core problem is that LLMs process instructions and data in the same input stream. A developer writes a system prompt to define the model’s behavior. A user sends a message. An external tool returns content for the model to process. From the model’s perspective, all of these are tokens in a sequence. It has no inherent mechanism for distinguishing a legitimate instruction from a malicious one embedded in retrieved content.
This architectural property is what makes prompt injection possible, and it cascades into nearly every other risk category OWASP GenAI Security Project GenAI Security Project GenAI Security Project identifies.
What is a large language model?
A large language model is an AI system trained on large amounts of text to understand and generate human language. It works by predicting the most likely next token based on its input, enabling it to produce coherent, contextually relevant text across a wide range of tasks.
What is the OWASP GenAI Security Project GenAI Security Project GenAI Security Project Top 10 for LLM Applications?
It is a framework published by OWASP GenAI Security Project GenAI Security Project GenAI Security Project that identifies the ten most critical security risks specific to applications built on large language models. The list covers prompt injection, insecure output handling, training data poisoning, excessive agency, supply chain vulnerabilities, and five additional risk categories. It is the primary reference for security teams assessing LLM deployments.
What is the difference between an LLM and a traditional AI model?
Traditional ML models are trained for specific, narrow tasks: classifying images, detecting spam, and predicting churn. LLMs are trained at scale on general text and can perform a wide range of language tasks without task-specific training. This generality is also what makes them harder to secure: their behavior is harder to predict and constrain than a purpose-built classifier.
What is indirect prompt injection?
Indirect prompt injection is an attack where malicious instructions are embedded in content that the LLM retrieves or processes on behalf of a user, rather than in the user’s direct input. A webpage, document, or API response containing hidden instructions can cause an LLM agent to take actions the user did not request and the developer did not intend.
How should organizations secure LLM-powered applications?
Start with the OWASP GenAI Security Project GenAI Security Project GenAI Security Project Top 10 for LLM Applications. Treat all LLM output as untrusted input to downstream systems. Apply least privilege to any agent given tools or external access—audit third-party models, plugins, and fine-tuning data sources before integration. Monitor model behavior in production for anomalies that may indicate prompt injection or policy violation.
