Technology glossary

Ablation Study

An experiment that removes or disables one component of a model at a time to measure how much that component contributes to performance. It can identify useful inputs, but it does not by itself prove causal relationships.

Absolute magnetometer

An instrument designed to determine an unambiguous magnetic-field value from a physical reference or response, rather than only tracking changes relative to a previously calibrated point.

Access Token

A credential represented by a string of data that a service accepts as proof that a client may perform specified actions or access specified resources. Tokens can be limited by scope and lifetime, but possession alone may be enough to use them, so issuance, storage and reuse must be controlled and logged.

Acoustic Embedding

A numerical representation learned from properties of an audio signal, such as rhythm, pitch and voice quality, that can be used by a machine-learning model.

Acoustic Identification

A method of identifying animals, machines, events, or environmental conditions by analyzing distinctive sound patterns in recorded audio.

Acrylic Acid

An industrial chemical used to make polymers, coatings, adhesives, and absorbent materials.

Action Chunking

Action chunking is a robotics and machine-learning technique that represents multiple low-level control steps as a compact action sequence, reducing how often a model must generate new commands while balancing speed against precision.

Action Validation

A checkpoint that evaluates a proposed automated action against policy, context and risk before the system executes it.

Actively Exploited Vulnerability

A security weakness for which reliable evidence shows that a malicious actor has exploited it in a system without the owner's permission. Under the EU Cyber Resilience Act, a manufacturer's awareness of active exploitation starts mandatory reporting deadlines.

Adaptation Limit

A point beyond which a person, ecosystem or system can no longer reduce climate risk to an acceptable level through available adaptation measures. Limits may be soft and changeable, or hard and effectively irreversible.

Adaptive Laboratory Evolution(ALE)

A method that repeatedly grows microorganisms under a chosen pressure so that variants better suited to the condition become more common over many generations.

Adaptive Optics

A system that measures optical distortion and changes mirrors or other components in real time to correct it. Astronomy and free-space communications use adaptive optics to compensate for turbulence in Earth’s atmosphere.

Additive Manufacturing in Rocketry

The use of layer-by-layer manufacturing, commonly called 3D printing, to produce rocket parts such as engines, injectors, and structural components. It can reduce part counts and production time while enabling complex shapes that are difficult to make conventionally.

Advanced Packaging

Techniques that combine multiple chips and high-speed connections into one package or system.

Adversary-in-the-Middle Phishing(AitM)

A phishing technique in which an attacker operates a real-time proxy between a victim and a legitimate sign-in service. The proxy relays the genuine authentication exchange while capturing credentials, multi-factor codes and authenticated session tokens.

Age Assurance

The set of methods an online service uses to estimate, verify or confirm whether a user falls within a required age range. It can include self-declaration, document checks, facial age estimation, account history or signals from third parties, with different levels of certainty and privacy impact.

Agent Audit Trail

A chronological record of actions an AI agent proposed, attempted and completed, including approvals and denials.

Agentic AI

An AI system designed to pursue an objective through multiple steps by selecting actions, using tools, observing results and revising its plan. Its autonomy depends on the permissions, tools, supervision and limits configured around the model.

Agentic Computing

A computing approach in which AI systems plan and carry out multi-step actions across software, tools, and services.

AI Accelerator

A specialized processor designed to perform the matrix calculations used in artificial intelligence training and inference more efficiently than a general-purpose CPU.

AI Agent

A software system that uses an AI model to choose and execute multiple steps toward a goal, often by calling tools and evaluating intermediate results.

AI Agent Sandbox

A restricted computing environment that limits an AI agent's access to files, networks, tools and credentials.

AI Capability Threshold

A defined level of measured performance or functional ability that triggers additional review, safeguards or legal obligations for an AI system.

AI Containment

A layered set of technical and operational controls designed to limit what an AI system can access, change or communicate during training, evaluation or deployment. Typical controls include sandboxing, restricted networking, least-privilege credentials, monitoring and emergency shutdown mechanisms.

AI Evaluation Environment

A controlled technical setting used to test an AI model or agent, including its prompts, tools, data, permissions, network access and monitoring.

AI Factory

A shared infrastructure and support program that combines high-performance computing, data services, software and expert assistance for developing and deploying AI systems. In the European program, AI factories provide startups, companies and researchers with access to publicly funded supercomputing resources.

AI Infrastructure

The data centers, power systems, networks, chips, cooling equipment, and physical facilities required to develop and operate AI systems.

AI model containment

Technical and operational controls that restrict what an AI model or agent can access and do during training, evaluation or deployment. Containment may combine sandboxes, virtual machines, network restrictions, monitoring and automatic stopping systems.

AI Music Generator

A generative system that creates audio, songs or musical components from text prompts, uploaded audio or other instructions. Such systems may produce melody, harmony, rhythm, instrumentation, vocals and lyrics, depending on their design and training data.

AI Post-Training

Training performed after a foundation model's main pretraining stage to shape its skills and behavior. It can include supervised fine-tuning, reinforcement learning and practice in tool-based environments. Post-training can improve coding, reasoning or safety, but may also strengthen unintended capabilities when related tasks appear in the data or reward environments.