Safety Glossary
Browse 47 safety terms defined in plain English, from the cultural dictionary of computing.
47 Safety Terms
- AI alignment
- The research discipline concerned with ensuring that artificial intelligence systems reliably pursue goals and behaviors that are beneficial to humans....
- AI Alignment Problem
- The challenge of ensuring advanced AI systems reliably pursue human goals, values, and constraints. It influences how models are trained, evaluated, or served,...
- AI Red Team
- A team that deliberately tests an AI system for harmful outputs, misuse paths, jailbreaks, and other safety or security weaknesses.
- AI Safety Benchmark
- A benchmark designed to measure model behavior on harmful, risky, deceptive, or policy-sensitive tasks.
- AI Sandbox
- A restricted environment used to test or execute AI-driven actions with limited permissions, controlled data access, and stronger safety guarantees. AI...
- Alignment
- Alignment in artificial intelligence refers to the challenge of ensuring that AI systems behave in ways that are consistent with human values, intentions, and...
- Canary Deployment
- A deployment strategy that gradually rolls out changes to a small percentage of users before deploying to the entire fleet. Named after canaries in coal mines...
- Code Hazard
- A piece of code or technical situation that creates material risk for correctness, security, or operability. In engineering slang, calling something a code...
- constitutional AI
- A training methodology developed by Anthropic where an AI system is guided by a set of written principles (a 'constitution') rather than relying solely on...
- Guardrail Culture
- A culture that relies on clear boundaries and safeguards to keep people and systems safe without requiring perfect judgment every time. Guardrail culture is...
- guardrails
- Safety mechanisms and filters applied to AI systems to prevent harmful, off-topic, or policy-violating outputs. Guardrails can be implemented at the prompt...
- Guardrails
- Validation layers applied to AI model inputs and outputs to enforce safety, accuracy, and compliance constraints. Can include content filters, format...
- Idempotency Key
- A unique identifier sent with an API request that allows the server to recognize and ignore duplicate requests, ensuring operations like payments are only...
- Immutable
- Data that cannot be changed after creation. Instead of modifying existing data, you create new copies with the changes applied. A cornerstone of functional...
- Interpretability
- The extent to which humans can understand why an AI system produced a particular output or how it represents information internally. Interpretability matters...
- Jailbreak AI
- An attempt to bypass an AI system's intended safeguards, restrictions, or instruction boundaries through adversarial prompting or workflow manipulation....
- Kill Switch
- Kill Switch is a built-in mechanism that disables functionality quickly under defined emergency or abuse conditions. Security teams use it to enforce trust,...
- Model Safety
- The set of behaviors, controls, and evaluation practices aimed at preventing harmful, unsafe, or policy-violating model outputs and actions. Model safety work...
- Moderation
- The process of identifying, filtering, or handling unsafe, disallowed, or harmful content in user inputs or AI outputs. Moderation is a core safety function in...
- Nullable
- A type annotation indicating that a variable may hold either a valid value or null/nil/None, requiring explicit handling before use in null-safe languages.
- Null Safety
- A type system feature that distinguishes nullable from non-nullable types at compile time, preventing null reference exceptions — the 'billion-dollar mistake'...
- Option Type
- A type that represents a value that may or may not exist, used to safely handle the absence of data without resorting to null references. Common variants...
- Personal Security
- The practices and protections an individual uses to secure their accounts, devices, data, and physical safety against digital threats such as phishing, account...
- Phantom Type
- A type parameter that appears in a type's definition but is not used in any of its fields — it exists only to carry compile-time information. Used to encode...
- Prompt Guard
- A safeguard or validation layer intended to protect prompts or detect dangerous prompt conditions such as injection, policy violations, or malformed context....
- Prompt Safety
- The practice of designing prompts and prompt-handling systems so they are resistant to misuse, injection, leakage, or unsafe behavior. Prompt safety overlaps...
- Red Team AI
- The practice of adversarially testing AI systems to uncover unsafe behavior, security weaknesses, misuse pathways, or policy failures before deployment. Red...
- Red Teaming AI
- Actively probing an AI system for unsafe behavior, jailbreaks, policy failures, or harmful outputs by simulating adversarial use.
- Refusal
- A response pattern in which an AI system declines to comply with a request because it is unsafe, disallowed, unsupported, or outside system boundaries. Good...
- Refusal Training
- Training or tuning methods intended to improve when and how an AI system refuses unsafe, disallowed, or unsupported requests. Refusal training is part of...
- Rollback
- Rollback is the process of reverting a system, database, or deployment to a previous known-good state after a failed update, bug, or incident. In database...
- Safety Alignment
- The work of shaping a model so its behavior stays within intended safety, policy, and usefulness boundaries in realistic use.
- Safety Benchmark
- A benchmark specifically designed to measure safety-related behavior such as refusal quality, harm avoidance, policy compliance, or robustness to adversarial...
- Safety Evaluation
- The broader process of assessing whether an AI system behaves safely under expected use and misuse scenarios. Safety evaluation may include red teaming,...
- Safety Policy
- A policy that defines what an AI system should and should not do from a safety and acceptable-use perspective. Safety policies guide training, prompting,...
- Safety Testing
- Testing focused on unsafe behavior, policy violations, prompt injection, abuse scenarios, and other safety-relevant failure modes. Safety testing is often done...
- Safety Training
- Training or tuning intended to improve safe behavior, refusal quality, and compliance with policies or human preferences. Safety training is one layer among...
- Sandbox Culture
- A team environment that encourages trying ideas in isolated, low-risk environments before exposing them to production or other critical systems. It reflects a...
- Sleeper Agent
- A hypothetical or experimental agent that appears benign under ordinary conditions but behaves differently when specific triggers are met. The concept is...
- Static Typing
- A type system where variable types are determined and checked at compile time rather than runtime, catching type errors before the program executes.
- Strong Typing
- A type system characteristic where implicit type conversions are restricted, requiring explicit casts and preventing operations between incompatible types,...
- Type System
- The set of rules a programming language uses to assign and check types, preventing certain classes of errors. Ranges from dynamic (Python, JS) to static (Rust,...
- Uncensored Model
- A model advertised as having fewer safety restrictions, moderation layers, or refusal behaviors than mainstream assistants.
- Union Type
- A type that represents a value which can be one of several specified types. Written as A | B in TypeScript, the compiler ensures you handle all possible...
- Unlearning
- Techniques for removing or suppressing specific information, behaviors, or data influence from a trained model without fully retraining from scratch.
- Watermarking AI
- Embedding detectable signals in model outputs or generated media so AI-produced content can later be identified.
- Zig
- A systems programming language positioned as a pragmatic alternative to C, with no hidden control flow, no hidden allocators, and no undefined behavior. Zig...
Related Topics
- Ai (19 terms in common)
- Type System (7 terms in common)
- Evaluation (5 terms in common)
- Security (4 terms in common)
- Deployment (3 terms in common)
- Alignment (3 terms in common)
- Language (3 terms in common)
- Training (2 terms in common)
- Prompts (2 terms in common)
- Privacy (2 terms in common)
- Ethics (2 terms in common)
- Typescript (1 terms in common)
- Systems (1 terms in common)
- Null (1 terms in common)
- Detection (1 terms in common)
- Policies (1 terms in common)
- Models (1 terms in common)
- Benchmarks (1 terms in common)
- Kotlin (1 terms in common)
- Open Models (1 terms in common)