Giskard

Giskard 3, an August 2026 ground-up rewrite of this open source library, targets LLM agents rather than the mix of tabular ML and LLM scanning in earlier releases. It comes as small packages: giskard-checks scripts test scenarios, including multi-turn chats, with assertions from regex and similarity matches up to LLM-as-judge checks for groundedness and policy conformity. giskard-scan is the red-team half: describe what your agent does in a sentence and it writes attack scenarios around that, trying prompt injection, jailbreaks, toxic or biased replies and false claims, while a separate quality scan grades RAG answers against a knowledge base. The target can be any sync or async Python function. It needs Python 3.12 or newer plus an API key for the LLM provider acting as judge, gpt-4o-mini by default. Usage analytics, minus your test content, are sent unless you opt out. The library is Apache 2.0 and runs locally; the company's paid Hub is a separate product. The older v2, with tabular scans and the RAGET generator, remains installable but unmaintained. A good fit for teams adding safety and regression gates to agent CI.