Loading page…
Simple definitions for a complicated conversation. Start here, then go deeper.
An AI system that can carry out a sequence of actions using tools or connected services.
Example: an assistant searches documents, prepares a draft and uses a tool to save it. Its authority depends on the tools and permissions granted.
See it in contextWork to make a system’s behavior and objectives fit intended human goals and values.
Example: an assistant optimizes for a quick booking when the user actually cares most about refundable travel. The intended goal needs clarification.
See it in contextWork to reduce risks from AI systems, including accidents, misuse, and unintended consequences.
Example: preventing a document assistant from exposing private records. The risk may arise from misuse, a mistake or a poorly designed workflow.
See it in contextProtecting AI systems, their data, and connected workflows against adversarial actions.
Example: preventing an attacker-controlled document from causing a disclosure. Security focuses on adversarial threats, though its controls can also reduce accidents.
See it in contextA structured test of specified behavior under defined conditions.
Example: reviewers score a fixed set of support conversations against a written rubric. The result applies to the recorded setup and test distribution.
See it in contextA risky case that a detector or evaluation fails to flag.
Example: a harmful response passes a check. A low false-alarm rate does not tell you how many risks the system missed.
See it in contextA harmless case incorrectly flagged as risky by a detector or evaluation.
Example: a moderation check flags a legitimate educational question. The label describes an error relative to the evaluation’s intended policy.
See it in contextThe extent to which a response is supported by the source material provided to the system.
Example: a summary follows the supplied report accurately. That does not establish that the report itself is correct or current.
See it in contextA control intended to constrain or check a system’s inputs, outputs, or actions. It may use a model or ordinary code.
Example: a content classifier and an enforced spending limit are both controls, but only the latter can directly reject an over-budget transaction.
See it in contextGenerated content that is incorrect or unsupported, even when it sounds plausible.
Example: an answer invents a citation. An unsupported claim may happen to be true; lack of support and demonstrated falsity are related but different findings.
See it in contextA person’s ability to understand, direct, review, or stop a system’s behavior.
Example: a reviewer can inspect recipients and attachments before sending, then reject or narrow the action. A notification after sending offers different control.
See it in contextAn attempt to circumvent a model’s intended behavior constraints.
Example: a user tries to persuade a model to disregard its behavioral constraints. An instruction attack can also arrive indirectly through external content.
See it in contextModel parameters that are available to obtain; actual usage rights depend on the license.
Example: a team downloads parameters to run a model locally. This does not automatically provide training data, unrestricted rights or a maintained application.
See it in contextPersonally identifiable information: data that can identify a person, alone or in combination.
Example: a name with contact details, or a combination of attributes that identifies someone. Exact definitions and handling obligations depend on context.
See it in contextAn attempt to get a system to treat untrusted content as instructions that override its intended task.
Example: a retrieved document tells an assistant to ignore the user and send data elsewhere. The document should supply information, not new authority.
See it in contextSystematic attempts to find failures and vulnerabilities through adversarial testing.
Example: testers intentionally vary document instructions to find a way around a tool boundary. Finding failures is different from estimating their frequency among ordinary users.
See it in contextThe surrounding controls, tool restrictions, checks, and monitoring used to manage a model in an application.
Example: a document agent combines restricted retrieval, tool approval, answer checks and incident logging. Test the combination, not just each component.
See it in contextAn isolated environment that restricts the resources and systems an operation can affect.
Example: generated code runs with a temporary folder and restricted network access. The actual restrictions matter more than calling the environment a sandbox.
See it in contextFurther context for these plain-language definitions. Examples illustrate the distinctions; terminology can vary between research and product documentation.
A framework for identifying, measuring and managing generative AI risks across the system lifecycle.
Threat examples and layered defenses for applications that process untrusted text.
Evaluates support for individual factual claims rather than treating a long answer as entirely right or wrong.
One explicit definition of openness, covering freedoms and access to more than model weights.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence