Skip to main content
SecureAI Guard

LLM security glossary

What each term actually means, where it comes from, and how it differs from the term it is most often confused with. Definitions are dated and cite the standard they come from.

  • AI bill of materials (AI BOM)

    An AI BOM is the inventory of models, datasets, weights and dependencies behind an AI system — what lets you answer whether you are affected after a disclosure.

    · 2 min read
    glossary
    supply chain
    governance

  • Context window

    The context window is the token budget a model attends to in one request. It is also the security boundary that does not exist: it is all one token stream.

    · 2 min read
    glossary
    fundamentals
    prompt injection

  • Data poisoning

    Data poisoning is the deliberate corruption of the data a model learns from — pre-training, fine-tuning or a retrieval index — to change its behaviour later.

    · 2 min read
    glossary
    data poisoning
    OWASP LLM Top 10

  • Denial of wallet

    Denial of wallet is an availability attack paid for in invoices rather than downtime. On metered model APIs the cost of a request is attacker-controlled.

    · 1 min read
    glossary
    availability
    cost

  • Embedding inversion

    Embedding inversion is reconstructing text from its vector. It is why a vector database must be classified at the same level as the documents it indexes.

    · 2 min read
    glossary
    RAG
    vector database

  • Hallucination

    A hallucination is fluent model output that is not true. It is a security concern, not just a quality one, once a downstream system acts on it.

    · 2 min read
    glossary
    misinformation
    OWASP LLM Top 10

  • Human in the loop

    Human in the loop means a person authorises an action before it takes effect. The most effective control against agent compromise, and the easiest to break.

    · 2 min read
    glossary
    agents
    controls

  • Jailbreak

    A jailbreak makes a model ignore its own safety training. It is related to prompt injection but not the same, and the difference decides who is harmed.

    · 2 min read
    glossary
    jailbreak
    prompt injection

  • Membership inference

    Membership inference is working out whether a record was in a model's training set. It leaks less than extraction, but membership can itself be sensitive.

    · 2 min read
    glossary
    privacy
    machine learning

  • System prompt leakage

    System prompt leakage is disclosure of the hidden instructions given to a model. The damage is rarely the wording — it is the secrets teams put beside it.

    · 2 min read
    glossary
    system prompt
    OWASP LLM Top 10

  • Training data extraction

    Training data extraction is recovering verbatim examples from a model's training set by prompting it. It turns a training decision into a disclosure risk.

    · 2 min read
    glossary
    privacy
    training data

  • Zero data retention

    Zero data retention means a provider does not persist your prompts or completions. The term means little until you pin down scope, duration and exceptions.

    · 2 min read
    glossary
    procurement
    data handling