Giskard is a continuous AI red teaming platform designed to protect AI agents. It helps users discover vulnerabilities that cause AI agents to fail before production, detect security risks, improve LLM security, and protect AI systems. AI agents are vulnerable to security attacks, which can pose risks to data and reputation; but overly secure agents may provide poor-quality responses to users. Giskard's goal is to find a balance between the two.
[Main capabilities]
- Continuous AI red teaming: continuously detect vulnerabilities, improve LLM security, and protect AI systems.
- Vulnerability detection methods: use internal knowledge (such as RAG knowledge bases) to identify hallucinations; detect issues such as stereotypes, discrimination, harmful content, personal information leakage, and prompt injection based on security vulnerability taxonomies; use external resources (cybersecurity monitoring and online data) to continuously identify new vulnerabilities; apply internal prompt templates (based on experience with various customers).
- Reporting and remediation: provide a structured report after each evaluation, giving a clear conclusion - deploy or fix; security vulnerabilities are ranked by severity, functional scenarios are tested according to requirements, and if passed, the Giskard Label is granted. From discovery to fix, Giskard handles the entire remediation management: qualify each finding, discuss severity with the team, prioritize, create tickets in the workflow, and rerun each test until the fix is confirmed.
[How to use]
- Pre-deployment use: provide comprehensive quantitative KPIs to ensure the AI agent is production-ready.
- Post-deployment use: continuously detect new vulnerabilities that may appear after the AI application is put into production.
- Supported types: Giskard Hub specifically supports conversational AI agents in text-to-text mode. Giskard operates as a black-box testing tool, and the Hub does not need to know the internal components of the LLM agent (foundation model, vector database, etc.). The entire bot only needs to be accessed through an API endpoint.
- Core advantages: automated vulnerability detection (not only testing AI, but also automatically detecting key vulnerabilities such as hallucinations and security flaws); proactive monitoring (detect AI vulnerabilities before users discover problems, continuously generating different attack scenarios and potential hallucinations); for business stakeholders (not only a developer tool, but also for domain experts and product managers, providing a collaborative red teaming playground and annotation tools).
[Applicable scenarios]
- Enterprise security teams, product teams, and domain experts who need to evaluate and ensure the security of conversational AI agents.
- Hallucination detection, factuality checks, and robustness testing for AI agents.
- Supervising AI Assistant behavior to prevent AI from providing inappropriate guidance to customers.
- Scenarios that require continuous testing of LLM agents and maintaining safety monitoring both before and after deployment.
- Organizations that need to manage AI vulnerabilities throughout the entire process from discovery to remediation.
[Limitations and Notes]
- Giskard Hub supports conversational AI agents in text-to-text mode; other types are not provided.
- Data residency and isolation: Optionally process data in the EU or US, benefiting from automatic updates and maintenance while protecting data.
- Fine-grained access and IP control: Supports role-based access control (RBAC), audit trails, and integration with identity providers; protects intellectual property through a zero-training policy.
- Compliance and security: End-to-end encryption at rest and in transit; as a European entity, provides native GDPR compliance, as well as SOC 2 Type II and HIPAA compliance.
- On-premises deployment: For mission-critical workloads in the public sector, defense, or other sensitive applications, the engineering team can help install Giskard Hub in an on-premises environment.
- Pricing: Specific pricing information is not currently provided; please refer to the official link.
- For a comparison of the features of Giskard Hub (Enterprise Edition) and Giskard Open-Source (Personal Edition), please refer to the official documentation.
- Differences from LLM platforms such as LangSmith: Giskard not only tests AI but also automatically detects critical vulnerabilities, uses internal and external data sources to automatically generate test cases, emphasizes proactive monitoring, and is designed for business stakeholders.
- Remediation support: After subscribing to Giskard Hub, you can optionally receive technical consulting support from the AI security team to help mitigate vulnerabilities and design effective guardrails in production environments.