Sources / Glossary

The references behind the field guide

Primary frameworks, maintained project documentation, safe training labs, and a plain-language glossary. External URLs are shown as visible, direct resources throughout every module.

Source document

Codelivly AI for Hackers - Beginner's Edition
Initial topic map supplied by the user; not reproduced.
Open source

Standard guidance

NIST Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
Current vocabulary for evasion, poisoning, privacy, and misuse attacks.
Open source

Risk framework

NIST AI RMF: Generative Artificial Intelligence Profile
Cross-sector guidance for mapping, measuring, and managing generative-AI risk.
Open source

Threat knowledge base

MITRE MITRE ATLAS
Living tactics, techniques, mitigations, and case studies for predictive, generative, and agentic AI.
Open source

Application-security baseline

OWASP GenAI Security Project OWASP Top 10 for LLM Applications 2025
The maintained 2025 list, replacing the older list shown in the source PDF.
Open source

Agent-security baseline

OWASP GenAI Security Project OWASP Top 10 for Agentic Applications
Risks and mitigations for systems that plan, call tools, retain memory, and act.
Open source

Testing methodology

OWASP OWASP AI Testing Guide
Open methodology for testing AI applications and model-related controls.
Open source

Training lab

ML security baseline

OWASP OWASP Machine Learning Security Top Ten
A draft, frequently updated overview of common machine-learning security risks.
Open source

Official documentation

Official tutorial

PyTorch Learn the Basics
Maintained beginner workflow for datasets, models, optimization, and model persistence.
Open source

Official security guidance

Hugging Face Hub security and pickle scanning
Why untrusted serialized models can execute code and how to reduce the risk.
Open source

Open-source tool

Cloud guidance

Microsoft Azure AI security best practices
Identity, network, model, data, monitoring, and red-team controls for Azure AI workloads.
Open source

Cloud framework

Google Secure AI Framework
A lifecycle view of secure AI design, deployment, and operations.
Open source

Secure-by-design guidance

CISA and UK NCSC Guidelines for Secure AI System Development
International guidance across secure design, development, deployment, and operation.
Open source

Security guidance

ENISA Securing Machine Learning Algorithms
Threats, vulnerabilities, and controls across the ML lifecycle.
Open source

Research practice

OpenAI Advancing red teaming with people and AI
Why manual, automated, and mixed red teaming complement one another.
Open source

Current research

OpenAI GPT-Red: Unlocking Self-Improvement for Robustness
A July 2026 example of automated prompt-injection red teaming and its limitations.
Open source

Related guide

Vanderburgh.it EU AI Act 2026 amendments guide
A practical summary of Regulation (EU) 2026/1744 and enterprise implications.
Open source

Plain language

AI security glossary

Use these definitions as a starting point, then follow the NIST and MITRE sources when precision matters.

Adversarial example: An input deliberately changed to cause a model to make a chosen mistake under a defined threat model.
Agent: An AI-enabled system that can pursue a goal through planning, tools, memory, and multi-step actions.
Attack surface: The reachable components and interfaces through which an adversary could influence confidentiality, integrity, availability, or safety.
Data poisoning: Manipulating training, adaptation, feedback, retrieval, or evaluation data to alter system behavior.
Embedding: A numeric representation used to compare or retrieve semantically related content; it is not an authorization or truth mechanism.
Evasion: Changing inference-time input so a model produces an incorrect or attacker-chosen result.
Guardrail: A policy or technical control intended to constrain inputs, outputs, actions, or use; effective systems use multiple layers.
Indirect prompt injection: Untrusted instructions delivered through content the model reads, such as a document, webpage, email, or tool response.
Inference: Running a trained model on input to produce a prediction, score, or generated output.
Invariant: A security property that must remain true, such as one user never retrieving another user's private documents.
Model extraction: Using access or queries to reproduce model behavior, parameters, or valuable capability.
Model provenance: Evidence showing where a model came from, how it changed, who approved it, and which version is deployed.
Prompt injection: Attacker-controlled content being interpreted as instructions that conflict with the application's intended policy.
RAG: Retrieval-augmented generation: selecting external content and adding it to model context before generating a response.
Red teaming: Structured adversarial testing intended to discover and understand system risks before wider harm occurs.
Threat model: A documented view of assets, adversaries, capabilities, trust boundaries, security goals, and plausible failure paths.
Tool calling: An application pattern where a model proposes or selects structured operations that external code may execute.

Editorial note

Original adaptation, not a reproduction

The user-provided 2026 PDF supplied the initial topic map. This field guide is an original rewrite with a defensive-first structure, current sources, safer labs, corrected terminology, and enterprise operating context.