Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, why it matters here, what it is, what it is not, and how it shows up in the insurance example.
Two sentences hold the words apart. A prompt asks. A guardrail enforces.
AbstentionThe agent, or a guard on its behalf, does not give a recommendation it should not give.Access control listA list attached to a resource that names who may read, change or delete it.Adversarial suiteThe attacks you already know, kept as tests and run against every change.AgentA program that uses a language model to decide its next step, calls tools to gather evidence or act, and repeats until it is done.AllowlistA list of permitted items, with refusal as the default for anything not on it.Approval fatiguePeople who approve many routine requests stop reading them, so the rare risky one passes too.Approval gateA place where an action does not take effect until an authorized person accepts it.Attempt number1 for the first try of a step inside a run, 2 for the first retry, and so on.Audit eventThe record of one guard decision, with the ids that tie it to the run, the step, the record, the rules and the evidence.Audit trailThe ordered, append-only set of audit events of a run, kept so the run can be reconstructed and checked later.Base rateHow common the thing you are looking for is in the traffic the check really sees.Blast radiusThe worst a component can do, computed from what it holds, assuming it is fooled.Break-glassA prepared emergency procedure that lets named people change or turn off a control for a short time, with an alarm and a record.Business ruleA rule about what the business accepts, such as which risks the insurer takes, written as code with an id.Bypass rateOf the attacks in the suite, how many got through to a harm.BypassThe attack worked, whichever guards fired along the way.CitationA pointer from a claim to its source, here the rule ids of the policy fixture.ClassifierA model that reads an input and returns a label or a score, such as spam 0.97.Coding agentAn agent that reads and edits files, runs shell commands and pushes branches in a repository.Confused deputyA program that holds a permission and uses it on behalf of a party that does not hold it.Context windowThe instructions, conversation, documents and tool results that a model reads before it writes its next output.ControlThe umbrella word: anything that lowers a risk, whether it advises, enforces, measures or assigns accountability.Covert channelA path that was never designed to carry data and can still signal it: timing, the choice between allowed calls, the number of calls, the length of a field.Decision logFor each decision: the input, the version, the rules applied, the result and an id that joins it to the request.Defense in depthPlacing more than one control in front of each harm, so the failure of one does not let it through.DelimiterTags or symbols placed around a piece of text, for example around each tool result.DeterministicA check is deterministic when it gives the same answer for the same input on every run.Direct injectionA prompt injection written by the person interacting with the agent.Egress allowlistA list of permitted outbound hosts; every other outbound connection is refused at the network edge.EgressAny connection from inside the system to somewhere outside it.Enforcement pointThe line in the request path where a check runs and its decision takes effect.EscalationMoving a case up to someone who may decide what the current level may not.EvaluationRunning the agent or a check over many cases and reporting how often the result was right.Evidence idAn id for a piece of evidence, such as one roof assessment, cited by claims and audit events.ExceptionA recorded permission to depart from one rule for one case, granted by a named role, with a reason, evidence and an expiry.Exfiltration pathA route by which data can leave the system for a place an attacker can read: links, images, tool arguments, error messages, requests to allowed hosts.ExfiltrationData sent out of the system to a destination chosen by an attacker.Fail closedThe mode in which a request is refused or held for a person when the check that should decide on it cannot run.Fail openThe mode in which a request passes as if allowed when the check that should decide on it cannot run.False negativeAn attack or harmful input that the check does not flag.False positiveHonest input treated as a problem: a false alarm.False-block rateOf the honest requests in the benign suite, how many were refused, held, denied a normal tool or ended differently than expected.GovernanceThe organization's decisions about ownership of rules, required controls and the path for exceptions.GroundingAn answer is grounded when each claim can be traced to evidence retrieved in this run.Guard decisionallow, deny, redact, flag or escalate: what one guard decided about one subject.GuardrailA condition that restricts or checks what an agent does, decided by code at a defined place, and holding whatever the model wrote.Guardrails AIA Python library that checks values going into and coming out of a model with reusable validators, each with its own action on failure.HookA script the harness runs at a fixed point in the agent's loop, with the proposed call as input; a hook before a tool call can block it.Human in the loopA design in which the system recommends and an authorized person decides; nothing takes effect until the person acts.human_decision{decided_by: role, decision, override, reason}, written only by a person's action.Idempotency keyA value the caller sends with a write so that the same write sent twice is stored once.Indirect injectionA prompt injection placed in a document, web page, tool result or retrieved passage, never sent to the agent by the attacker.Input checkCode that checks or edits a request before any model sees it.InstructionA sentence in the model's input that asks it to do or avoid something. The model usually follows it; nothing makes it.IsolationArranging a system so that one tenant or one run cannot see or affect another.JailbreakAn attempt to get a model to produce what its maker trained it to refuse.LeakagePersonal data, contacts, secrets or another tenant's ids showing up in an answer.Least privilegeGive an identity exactly the permissions its task needs, and nothing it might need someday.MCPAn open protocol by which an AI application connects to servers that offer tools, data and prompt templates, using JSON-RPC messages.MCP serverA service, often written by another party, that offers tools or data to an AI application over the Model Context Protocol.MiddlewareCode that runs before (and sometimes after) every handler, so a check cannot be forgotten on a new endpoint.Model-based guardrailA guard that asks a model or classifier how much an input looks like a problem, and acts when the score reaches a threshold.Multi-agent systemA system in which several agents, each with its own instructions and tools, send work and text to each other.NeMo GuardrailsA runtime that sits between an application and the model and runs configured checks, called rails, at the input, retrieval, dialog, tool-call and output stages.NIST AI RMFA document from the US standards institute NIST that describes how an organization manages risks of AI systems, in four functions: Govern, Map, Measure, Manage.NormalizationTransforming text so that variants of the same thing look the same to a check.Open Policy AgentA general-purpose engine that takes structured input and returns a decision from policies written in Rego.Output checkCode that checks the finished answer before it is stored, shown or acted on.OverrideThe person decides differently from what the agent recommended.OWASP Top 10 for LLM ApplicationsA document that names the ten most important risks of applications built on language models and the mitigations for each. The current edition is 2025.Parameter boundsRules on a call's arguments: which ids, which hosts, which ranges are allowed.PayloadThe instruction inside an attack text, as opposed to the framing around it.PermissionWhat an identity is allowed to do, independent of what it is asked to do.Personal dataData that identifies a person alone or with other data, such as an identifier number or a private email address.Policy as codeThe rules a system must follow, written so a program evaluates them, each with an id, grouped in versions.Policy complianceThe answer is one the rules allow for these facts under the pinned policy version.Policy conflictA case to which two rules, or two checks, both apply and give different answers.Policy decision pointThe component that takes facts and returns a decision; everything else asks it instead of deciding.Policy engineA separate component that takes a question and facts and returns a decision, so rules live apart from the application code.Policy fixtureThe invented underwriting rules every page uses: R-ROOF-01, R-ROOF-02, R-ROOF-03, R-SPRK-01, R-HUMAN-01.Policy versionThe version of the rules a decision was made under, such as 2026.2.PrecedenceThe fixed order that decides which rule or check wins when several apply and disagree.PrecisionWhen the check raises its hand, how often it is right.ProbabilisticA component is probabilistic when it samples its output, so one test does not tell you what the next run will do.Prompt injectionText in the model's input, written by someone else, that the model acts on as if the people who run the agent had written it.PromptfooA command-line tool and library that runs prepared and generated attacks and tests against an application and reports which ones got through.ProvenanceThe record of the source of a piece of data and the hands it passed through.Provider safety toolA safety feature that a model provider offers next to its model: a classifier endpoint such as moderation, or published guidance on defending an application.RAGBefore the model answers, the program searches a store of documents and pastes the passages it found into the model's input.RecallOf the attacks that arrive, how many this check sees.RecommendationA proposal to the underwriter: approve, decline, request_info or escalate, with reasons, cited rule ids and a policy version.Red teamDefenders acting as attackers against their own system, to find failures before anyone else does.RedactionRemoving a piece of sensitive text and putting a marker in its place.RegoThe language in which OPA policies are written as rules over input and data.Regression setEvery incident and red-team finding, turned into a synthetic test that runs on every change.Role-based access controlInstead of listing people on every resource, give permissions to a role and give people the role.Run idThe id of one execution of the workflow for one submission.SandboxAn environment that limits what running code can read, write and connect to.Schema validationComparing a record with its schema: which fields must exist, their types, and the values allowed.ScopeA label in an identity's credentials that names one permission, such as recommendation:write.Shadow modeRunning a new or changed check on live traffic so that it records what it would have decided while its decision changes nothing.SubmissionThe broker's email turned into one JSON record with the facts of the request.System promptThe text the application puts in front of every exchange with the model: the role, the task and the instructions.Taint trackingThe mechanism that carries the label of untrusted text through every copy, so that anything derived from it is untrusted too.TaintA mark saying 'this came from untrusted data' that is carried along when the data is copied, summarized or combined.TenancyThe rule that many customers share one system and each may act only on its own records.ThresholdThe line that turns a score into a decision: at or above it the check flags or blocks.Tool callA structured request from the model: a function name and arguments the model chose. The program decides whether to run it.Tool poisoningThis site's name for tool metadata, such as a description, written to give the model instructions instead of describing the tool.Tool-output launderingUntrusted text, usually from a tool result, coming back later as the agent's own notes, a summary, a cache entry or a message, so that it looks trusted.ToxicityHarmful language in a model's input or output, usually detected by a classifier with categories and scores.Trace idA 32-hex-character id shared by every span (step) recorded for one run.Trust boundaryA line in a system where data from outside meets data from inside, so it must not be trusted the same way.Trust labelA small record, written by code, that says where a piece of context came from and whether the insurer controls it.Underwriter cockpitThe screen that shows a recommendation and records the underwriter's decision.UnderwriterThe person at the insurer with the authority to decide on a submission.Underwriting agentThe agent that reads a submission, gathers evidence and writes a recommendation for the underwriter.Untrusted inputText the system owner does not control, whatever its author intends.ValidationCode that checks a value is acceptable as data before any logic runs, and refuses it otherwise.
Each term is explained in full on one page and used everywhere else. Where the words come from is on About.