AI Glossary

An AI system that can take independent actions in the real world or digital environments without constant human direction. Can make decisions and execute multi-step tasks.

A standardized test or evaluation used to measure and compare the performance of AI models across specific capabilities or tasks.

An incident where an AI system escapes its intended constraints or safety boundaries and behaves unexpectedly.

The process of further training a pre-trained AI model on specific data to adapt it for particular tasks or domains.

The most advanced AI models currently available, representing the cutting edge of AI capability. Examples include GPT-5.6, Claude 3.5, and similar state-of-the-art large language models.

Safety mechanisms built into AI models through training and architecture to prevent dangerous outputs. These include refusal to help with illegal activities, violence, or other harmful tasks.

When an AI model generates false, misleading, or fabricated information while sounding confident, rather than admitting uncertainty.

The process of running a trained AI model to generate outputs based on input data, as opposed to the training phase.

Techniques used to circumvent a model's safety constraints and guardrails, causing it to behave in unintended ways or produce harmful outputs.

The time it takes for an AI model to process an input and generate an output. Lower latency means faster response.

An AI model whose weights (parameters) are publicly released, allowing anyone to download and run the model locally. Examples include Llama, Mistral, and Bloom.

A database security mechanism that controls which rows of data different users can access.

An isolated computing environment used for testing AI models safely. It restricts the model's access to the internet, file systems, and other resources to prevent unintended consequences.

The basic unit of text that an AI model processes. One token roughly equals 4 characters or a word in English.

A previously unknown security flaw in software that developers are unaware of, making it exploitable before a patch can be released.