Glossary

Terms used in Quilr Red Teaming and in these docs.

Term	Definition
Assessment	A configured test of one target: its mode, coverage, custom attacks, and schedule.
Run	One execution of an assessment. Runs are listed under Recent runs.
Mode	The kind of target being tested: AI App / Agent, AI Workflow, or Model.
Objective	One attack goal, such as System Prompt Extraction. The attack library holds the built-in objectives.
Quick scan	About 8 core objectives, for a fast check.
Full library	Every objective in the attack library.
Custom objective	An objective you write for your own target, reported under a Custom category.
Synthesized	An objective generated for this specific target, often to test its tools.
Finding	A weakness a run found, with severity, evidence, a recommended fix, and framework mappings.
Breached / Partial / Held / Inconclusive	The outcome of an attack. See Risk grade and severity.
Pass rate	The share of attacks the target withstood.
Risk grade and risk score	A letter grade from A to F and a score out of 100 for a run. Lower risk is better.
Evaluation mode	How results are judged. Hybrid combines deterministic evidence with model judges; Deterministic uses evidence detectors only; Judge uses model judges only.
Judge consensus	How strongly the model judges agreed on a result.
Replay-safe	A target where repeating an attack cannot cause a harmful or duplicate side effect. Confirmation replays run only on replay-safe targets.
Replay	Sending a finding's attacks again, usually after a fix, to check whether it still works.
Authorized scope	The record of what you are approved to test: the target, a ticket or reference, and the testing boundary.
Detect	Automatic discovery of how a target is reached and what request and response format it uses.
Via Direct	The attack technique that succeeded, shown on each finding in the Full report.