AI Foundations
This learning method involves a “State → Action → Reward” loop repeated thousands of times to teach an agent how to decide actions on its own
What is Reinforcement Learning?
The term for realistic audio or video impersonations generated by AI for fraud and deception
What are Deepfakes (or Voice Cloning)
This guardrail type is used to automatically mask or remove sensitive information like names, addresses, or credit card numbers
What is a PII/PCI Redaction Guardrail?
These are the three essential elements of a high-quality user prompt required to maximize model capability
What are Context, Perspective, and Output Shape?
This role is responsible for building and managing data pipelines to ensure high-quality data is accessible for AI systems
Who is a Data Engineer?
This specific AI type produces original content by learning from large datasets and mirroring the patterns found within them
What is Generative AI?
This threat occurs when an attacker manipulates instructions to bypass a model's safety rules or override active system instructions
What is Prompt Injection?
The security best practice that ensures users and subjects receive only the minimum access necessary for their specific role
What is Least Privilege?
These usage limitations control which specific AI input/output types—such as text, images, or audio—are permitted by the system
What are Modality Limits?
This voluntary U.S. framework organizes AI risk management into four functions: Govern, Map, Measure, and Manage
What is the NIST AI Risk Management Framework (RMF)?
This data type carries the highest security risk because it often contains hidden sensitive data throughout its large volume
What is Unstructured Data?
This component of the MIT AI Risk Repository explains how, when, and why AI risks occur by classifying them by entity, intent, and timing
What is the Causal Taxonomy?
This data safety measure substitutes real data with generic placeholder labels while protecting the format of the original data
What is Data Masking?
The data safety measure used when sharing reports that may contain sensitive internal details that must not be visible or recoverable
What is Data Redaction?
This regulation categorizes AI applications into risk tiers, including "High risk" and prohibited "Unacceptable risk"
What is the EU AI Act?
This AI technology uses multi-layered neural networks to learn complex patterns from large amounts of raw data and adapt as new threats emerge
What is Deep Learning?
This attack involves an adversary using repeated queries to learn confidential information about an AI model or to recreate it entirely
What is Model Extraction?
These isolated hardware environments provide secure processing for "data in use" to prevent unauthorized access to sensitive information
What are Trusted Execution Environments (TEEs)?
This testing method, often used in SOC operations, identifies bugs, errors, and inconsistencies within source code
What is Code Linting?
In this training approach, the model is brought to a local device and trained with user data so that the raw data never leaves the device
What is Federated Learning?
This 7-layer framework structures AI risks across layers such as Foundation Models, Data Operations, and the Agent Ecosystem
What is the MAESTRO framework?
This encryption method allows computation on encrypted data without decrypting it, keeping the data protected while it is being processed
What is Homomorphic Encryption?
This practice integrates software development, security, and operations throughout the CI/CD pipeline
What is DevSecOps?
This principle of Responsible AI ensures that stakeholders can understand the reasoning behind a system’s output
What is Transparency and Explainability?