AI security & deployment glossary

All terms

• AI governance

  The policies, controls and accountability structures that ensure AI systems are used safely, lawfully and in line with an organisation’s obligations.

• AI red-teaming

  Structured adversarial testing of an AI system to find how it can be made to fail, leak data or be misused before real attackers do.

• Air-gapped AI

  An AI system that runs on hardware with no network connection to the internet or any untrusted network, so data cannot leave the environment.

• Data residency

  The requirement that data is physically stored and processed within a specified geographic location, such as a particular country or region.

• Data sovereignty

  The principle that data is subject to the laws and governance of the jurisdiction in which it is collected, stored or processed.

• Fine-tuning

  Further training of a pre-trained model on additional, task-specific data to adapt its behaviour to a particular domain or use.

• Inference

  The process of running a trained model on new inputs to produce outputs — using the model, as opposed to training it.

• Model exfiltration

  The unauthorised extraction of an AI model — its weights, architecture or behaviour — or of the sensitive data it was trained on or has access to.

• Model risk

  The risk of adverse outcomes from decisions based on a model that is incorrect, misused or misunderstood.

• Model weights

  The numerical parameters a model learns during training; they encode what the model knows and constitute the model itself.

• OFFICIAL-SENSITIVE

  A UK government handling marking applied to OFFICIAL information that needs additional care because its loss could cause more serious harm.

• On-premises LLM

  A large language model deployed on infrastructure the organisation controls, so data is processed inside its own security perimeter rather than a vendor cloud.

• Open-weight model

  An AI model whose trained parameters are publicly released, so it can be downloaded and run on infrastructure the user controls.

• Private cloud AI

  AI deployed in a dedicated, single-tenant cloud environment the organisation controls, rather than a shared multi-tenant public AI service.

• Prompt injection

  An attack that hides malicious instructions in text an AI model reads, causing it to ignore its intended task and follow the attacker instead.

• Retrieval-augmented generation (RAG)

  A technique where a language model retrieves relevant documents at query time and uses them to ground its answer, rather than relying only on training.

• Sovereign AI

  AI capability deployed so that the data, models and infrastructure remain under the control and jurisdiction of a specific organisation or nation.

• Zero-egress deployment

  A deployment configured so that no data can leave the environment — outbound network paths are removed or blocked by default.