CerbIA is a Python library and CLI for evaluating AI agents' inputs and outputs. CerbIA composes security gates around AI agents' inputs and outputs. Configure loaders, preprocessors, scanners, and score aggregators to evaluate content and return an actionable verdict. Use it in applications, locally, or in CI/CD to inspect prompts, files, and other agent interactions.
The same gate definition can be reviewed as YAML and reused across the command line and application code. A loader extracts input, configured preprocessors normalize it, scanners report findings, and a score aggregator combines those findings. See Architecture for the processing model and the prerequisites before trying the quickstart.
Start with Getting started, then use Configuration to define a pipeline. The package reference documents the public Python namespaces; optional capabilities are available through the CLI, ML, Presidio, and ProtectAI packages.
Key features
-
Composable evaluation pipeline: combine loading, preprocessing, inspection, and scoring; see Architecture.
-
YAML gate configuration: keep gate definitions reviewable and reusable; start with Configuration.
-
Local and automated checks: run the same gate from the CLI or an application; see CLI reference.
-
Task-based walkthroughs: move from installation to a working scan in Guides.
-
Optional detection integrations: add local ML, Presidio, or ProtectAI packages when needed; see Package modules.
Scope
Scanner results are heuristic or model-based signals, not a security certification or a replacement for application controls. Evaluate each gate against your threat model and retain defense in depth. CerbIA provides reusable inspection components; it does not make a universal security verdict for every application. For exact configuration and command details, use the Reference section rather than treating this overview as an exhaustive catalog.