Giskard

Giskard is an open-source Python library for testing AI agents, paired with a commercial security platform and assessment service. It's for developers checking agent behavior and security teams assessing deployment risks. The library runs in your own environment under Apache 2.0; Giskard Hub is available hosted or on-premise.

The library combines evaluations with automated attack generation. Giskard Checks supports ordinary assertions, LLM-based judges and multi-turn conversations, so tests can assess an interaction across several exchanges. Giskard Scan generates adversarial tests from a plain-language description of an agent, covering prompt injection, harmful content, stereotypes and misinformation. Teams can add their own test generators and evaluate RAG answers against a knowledge base for groundedness and quality.

Running the library locally doesn't make every test offline. LLM judges and scan generators require a model provider and API key; the documented integrations include OpenAI and Anthropic, so those evaluations send requests to the selected provider. Aggregated analytics are optional.

Hub tests text-based conversational agents through an API without needing access to their internal models or vector databases. It checks for data disclosure and security attacks alongside hallucinations, contradictions, omissions and incorrect refusals. The assessment service produces severity-ranked findings and a deploy-or-fix verdict, then helps prioritize fixes and retests them. Hosted processing offers EU or US data residency, while on-premise deployment supports workloads whose data must stay within the organization's environment.

The current v3 library replaces the earlier agent evaluation and RAG test-generation tools. Legacy v2 remains available but is no longer actively maintained, including its automatic tabular-model scan.