Exam glossary
GIAC GAIPS glossary
The terms you meet in the GIAC GAIPS exam, with a short definition. The exam is in English: the German equivalent helps if you study in German. Portuguese, Spanish and French will follow.
29 terms · Definitions checked against official documentation on November 13, 2026 · Auf Deutsch
A
- AI agent#German: KI-Agent
- An LLM system that plans steps and calls tools to reach a goal. Its security depends on which tools and permissions it has and on checks outside the model.
- Exam domain: Agentic Systems and AI Integrations
- AI supply chain#German: KI-Lieferkette
- Third-party models, adapters such as LoRA, datasets, libraries and plugins an AI system depends on. Each can be tampered with, so provenance and integrity checks matter.
- Exam domain: Development Pipelines and MLOps Security
- Alignment#German: Alignment (Ausrichtung)
- Training a model to follow intended values and rules, for example with reinforcement learning from human feedback (RLHF).
- Exam domain: Model Customization and Alignment
C
- Context window#German: Kontextfenster
- Everything the model sees in one request: system prompt, conversation, retrieved documents and tool results. Anything in it can influence the output.
- Exam domain: AI and LLM Foundations
D
- Data poisoning#German: Datenvergiftung (Data Poisoning)
- Manipulating training, fine-tuning or retrieval data to introduce backdoors, bias or wrong behaviour.
- Exam domain: Development Pipelines and MLOps Security
E
- Embedding#German: Embedding (Vektordarstellung)
- A numeric vector that represents the meaning of text, used to find similar content in a vector database.
- Exam domain: Knowledge Augmentation and Retrieval
- Evasion attack#German: Evasion-Angriff (Umgehungsangriff)
- Inputs crafted at inference time so the model misclassifies or misbehaves, one of the attack classes in NIST AI 100-2.
- Exam domain: AI and LLM Foundations
- Excessive agency#German: Übermäßige Handlungsbefugnis
- An LLM system with more functions, permissions or autonomy than it needs, so a manipulated output can cause real damage. Countered by least privilege and human approval for high-impact actions.
- Exam domain: Agentic Systems and AI Integrations
F
- Fine-tuning#German: Fine-Tuning (Nachtraining)
- Further training a pretrained model on specific data to change its behaviour. It can weaken earlier safety alignment and does not on its own stop prompt injection.
- Exam domain: Model Customization and Alignment
G
- Guardrails#German: Guardrails (Schutzschranken)
- Checks before and after the model, such as input filters, content moderation and output validation. They reduce risk but do not replace least privilege.
- Exam domain: AI Application Architecture and Development Frameworks
I
- Improper output handling#German: Unsichere Ausgabeverarbeitung
- Passing model output to browsers, shells, databases or other systems without validation or encoding, which can lead to XSS, injection or code execution.
- Exam domain: AI Application Architecture and Development Frameworks
J
- Jailbreak#German: Jailbreak
- A form of prompt injection that makes the model ignore its safety rules altogether.
- Exam domain: AI and LLM Foundations
L
- Large language model(LLM)#German: Großes Sprachmodell (LLM)
- A neural network trained on large amounts of text to predict the next token. It follows instructions in its context window but cannot reliably tell trusted instructions from untrusted data.
- Exam domain: AI and LLM Foundations
- Least privilege#German: Prinzip der minimalen Rechte
- Giving models, agents and service accounts only the permissions they need. The key control against excessive agency.
- Exam domain: AI Infrastructure and Deployment Security
- Low-rank adaptation(LoRA)#German: Low-Rank Adaptation (LoRA)
- A parameter-efficient fine-tuning method that trains small adapter weights instead of the whole model. Adapters from third parties are part of the supply chain.
- Exam domain: Model Customization and Alignment
M
- Membership inference#German: Membership Inference (Zugehörigkeitsangriff)
- A privacy attack that tests whether a specific record was part of the training data.
- Exam domain: AI Risk Management and Strategic Application
- MLOps#German: MLOps
- Practices for building, deploying and running machine learning models reliably: data pipelines, training, model registry, deployment and monitoring. MLSecOps adds security to each step.
- Exam domain: Development Pipelines and MLOps Security
- Model Context Protocol(MCP)#German: Model Context Protocol (MCP)
- An open protocol that connects AI applications to tools and data sources through MCP servers. Untrusted servers and tool descriptions are an injection and supply-chain risk.
- Exam domain: Agentic Systems and AI Integrations
- Model extraction#German: Modell-Extraktion
- Copying a model's behaviour or parameters by sending many queries and learning from the answers.
- Exam domain: AI Infrastructure and Deployment Security
- Model serialization#German: Modell-Serialisierung
- Saving a model to a file. Pickle-based formats can execute code when loaded, so untrusted models should use safer formats such as safetensors and be scanned and signed.
- Exam domain: Development Pipelines and MLOps Security
N
- NIST AI Risk Management Framework(AI RMF)#German: NIST AI Risk Management Framework (AI RMF)
- A voluntary framework for managing AI risks, organised in four functions: Govern, Map, Measure and Manage.
- Exam domain: AI Risk Management and Strategic Application
O
- OWASP Top 10 for LLM Applications#German: OWASP Top 10 für LLM-Anwendungen
- OWASP's list of the most critical risks for LLM applications; the 2025 edition runs from prompt injection (LLM01) to unbounded consumption (LLM10).
- Exam domain: AI Risk Management and Strategic Application
P
- Prompt injection#German: Prompt Injection
- Input that changes the model's behaviour against the developer's intent. Direct injection comes from the user; indirect injection hides in content the model processes, such as web pages, emails or retrieved documents. OWASP lists it as LLM01:2025.
- Exam domain: Agentic Systems and AI Integrations
R
- Retrieval-augmented generation(RAG)#German: Retrieval-Augmented Generation (RAG)
- The application retrieves relevant documents, usually from a vector database, and adds them to the prompt. Retrieved content is untrusted input and can carry indirect prompt injection.
- Exam domain: Knowledge Augmentation and Retrieval
S
- Sensitive information disclosure#German: Offenlegung sensibler Informationen
- The model or application reveals personal data, credentials or confidential business data, for example from training data, context or connected systems.
- Exam domain: AI Risk Management and Strategic Application
- System prompt leakage#German: Offenlegung des Systemprompts
- The risk that instructions in the system prompt are revealed. OWASP's point: the system prompt must not contain secrets or act as an access control.
- Exam domain: AI Application Architecture and Development Frameworks
T
- Tool calling#German: Tool Calling (Werkzeugaufruf)
- The model asks the application to run a function, such as an API call or database query. Authorisation for the call belongs in the application, not in the prompt.
- Exam domain: Agentic Systems and AI Integrations
U
- Unbounded consumption#German: Unbegrenzter Ressourcenverbrauch
- Uncontrolled use of model resources that leads to denial of service, runaway costs or model extraction. Countered with rate limits, quotas and monitoring.
- Exam domain: AI Infrastructure and Deployment Security
V
- Vector database#German: Vektordatenbank
- A database that stores embeddings and returns the nearest matches. Its risks include missing access control per document, poisoned entries and leakage between tenants.
- Exam domain: Knowledge Augmentation and Retrieval