Giskard

Giskard is an AI security and evaluation platform for conversational LLM agents. It tests agents as black boxes through an API endpoint, without requiring access to their foundation model, vector database, or other internal components. The platform is designed for both pre-deployment assessment and continuous testing after release, covering security risks and response-quality problems.

Its security capabilities include automated adversarial testing, multi-turn attacks, prompt-injection testing, data-disclosure checks, harmful-content and stereotype detection, and tool-calling validation. The platform also supports scenario-based tests that model specific personas and business logic. Giskard Guards adds a runtime control layer that evaluates prompts, tool calls, parameters, permissions, and responses against context-aware, policy-driven rules.

For quality evaluation, Giskard can generate domain-specific question-and-expectation pairs from a knowledge base, apply RAG quality metrics, and support custom evaluation metrics defined around business requirements. Human reviewers can inspect, debug, annotate, and prioritize results. Dataset import, tags, task assignment, audit trails, scheduled alerts, and CI/CD support connect testing with ongoing engineering and review processes.

After an assessment, the platform provides a structured report with a deployment verdict, ranked security findings, and functional scenarios tested against the team’s requirements. Giskard describes a remediation workflow that includes qualifying findings, discussing severity, prioritizing actions, opening tickets, and rerunning tests to confirm fixes. Deployment options described on the site include SaaS, private cloud, on-premises, and air-gapped environments, with enterprise controls for access, data residency, and support.