What Giskard does
Giskard provides an open-source LLM testing library and an enterprise Hub for adaptive vulnerability scans, datasets, annotations, and continuous red teaming.
Giskard concentrates on finding security and business-logic failures in LLM applications. The open-source library can generate tailored scenarios locally, while Giskard Hub adds an enterprise interface, broader probe catalog, collaboration, continuous red teaming, and governance. Current documentation emphasizes text agents, OWASP-aligned categories, multi-turn attacks, and converting findings into tests.
The open-source library is free under its license, but the chosen generation and judge models, compute, engineering, and remediation are not. Hub pricing is custom and should be scoped by agents, scans, probes, seats, environments, retention, integrations, support, and private deployment. A free trial may be offered, but it is not a universal production price.
Automated attacks can create harmful text, send agent descriptions and responses to the selected model provider, miss novel vulnerabilities, or produce false alarms. Run against sandboxed tools and synthetic records, cap cost and duration, disclose testing to owners, restrict transcripts, and never let a test agent send messages, change production data, or spend money. Security experts and domain owners must reproduce and triage findings.
How Giskard works
A team wraps an agent endpoint or function and describes its purpose, tools, and expected behavior. Giskard's open-source library uses a configured model to generate adversarial scenarios and another evaluation step to judge responses; Hub adds specialized multi-turn probes, security grades, datasets, task assignment, annotations, access control, and continuous scans. Reviewers inspect attack transcripts, confirm vulnerabilities, prioritize remediation, and rerun accepted cases as regressions.
How to set up Giskard
Authorize and scope the test
Name the agent owner, environments, allowed attack categories, sensitive data, tools, rate limits, notification plan, evidence handling, and stop conditions.
Build an isolated target
Use test accounts, synthetic data, non-production credentials, reversible tools, outbound-network restrictions, and explicit blocks on consequential actions.
Configure models and data flow
Choose generation and judge providers, review what they receive, protect keys, set languages and budgets, and document retention and deletion.
Run a bounded scan
Start with selected categories, monitor activity, inspect every high-severity transcript, reproduce likely failures, and separate product defects from judge errors.
Remediate and regress
Convert confirmed findings into owned tests, fix architecture and permissions rather than only prompts, rerun after changes, and schedule controlled reassessment.
Giskard FAQs
Is Giskard open source?
Its LLM testing library is open source. Giskard Hub is a separate enterprise platform with custom commercial terms.
What does a Giskard scan do?
It generates adversarial single- or multi-turn scenarios for an agent and evaluates responses for configured security and business failures.
Does it test every OWASP risk?
Coverage depends on edition, version, probe, agent interface, and configuration. A scan is one layer of a broader security program.
What data reaches model providers?
The open-source scan documentation says the configured models receive the agent description and responses; review each provider and avoid sensitive test data.
Can the security grade replace expert review?
No. Grades summarize automated findings. Experts must reproduce issues, assess exploitability and impact, and verify remediation.
Listing reviewed 2026-07-31. Product details and pricing can change; verify important terms on the provider's website.
Related Coding AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.