giskard-oss: Test and Red-Team LLM Agents

Overview

Giskard is a modular Python toolkit for testing and evaluating LLM-based agents, including multi-turn systems and RAG applications. It addresses the difficulty of validating systems whose outputs can vary, combining scenario-based checks with automated vulnerability scanning.

Use it to define evaluations for a callable system, run checks against its outputs or traces, and generate adversarial test suites. The current v3 packages focus on agent evaluation and security; legacy automatic tabular-model scanning remains in v2.

Key Features

• Define scenarios and suites for synchronous or asynchronous targets.
• Run checks such as string matching, regex, semantic similarity, groundedness, and LLM-as-judge assessments.
• Evaluate multi-turn interactions, not only individual prompts and responses.
• Scan agents with generated adversarial tests for prompt injection, harmful content, stereotypes, and misinformation.
• Evaluate RAG answer quality and grounding.
• Use modular packages for checks, scanning, shared utilities, LLM routing, and agent orchestration.

Use Cases

• Agent developers can add regression checks when prompts, models, or workflows change.
• RAG teams can test whether generated answers are grounded in retrieved context and meet quality expectations.
• AI safety and security teams can probe an agent for prompt-injection or harmful-output vulnerabilities before deployment.
• Teams with multi-step assistants can evaluate complete conversations and trace-dependent behavior.

Project Facts

• Language: Python
• License: Apache-2.0
• Stars: 5.9k
• Forks: 547
• Topics: agent-evaluation, ai-red-team, ai-security, ai-testing, fairness-ai, llm, llm-eval, llm-evaluation, llm-security, llmops, ml-testing, ml-validation, mlops, rag-evaluation, red-team-tools, responsible-ai, trustworthy-ai
• Archived: false

Getting Started

Requires Python 3.12 or newer. Install the base package, or add the scan extra for vulnerability and quality scanning:

pip install giskard
pip install "giskard[scan]"

See the repository README and documentation for configuration and examples.

Alternatives

• AgentEvals: AgentEvals focuses on evaluating agent trajectories with ready-made evaluators, while Giskard also targets broader AI security and vulnerability testing.
• Agent Anvil: Agent Anvil centers on CI scenario suites and failure analysis for tool-using agents, while Giskard also probes AI systems for security vulnerabilities.
• Promptfoo: Promptfoo uses declarative configs for evaluating and red-teaming prompts, agents, and RAG apps, while Giskard provides a Python toolkit for repeatable checks.
• EasyJailbreak: EasyJailbreak focuses on assembling and benchmarking jailbreak attacks, while Giskard covers broader agent evaluation and AI vulnerability testing.

Considerations

• LLM-as-judge checks and scan generators require a configured provider SDK and API key.
• The current v3 release is a rewrite focused on dynamic, multi-turn agent testing and has a Python 3.12+ requirement.
• Giskard v2 remains available but is no longer actively maintained. Automatic tabular/ML scanning is v2-only and is not planned for v3.
• Optional aggregated telemetry is described in the project documentation. The README provides environment variables to opt out.