Giskard

Automated red-teaming for conversational agents. You point Giskard at an agent's API and it runs multi-turn adversarial attacks against it — prompt injection, data disclosure, jailbreaks, hallucination, inappropriate refusals — and reports both the security failures and the business-compliance ones. Sold as the Giskard Hub (enterprise), Guards (runtime guardrails), and an open-source scanner. The research arm is unusually public: three named public benchmarks and datasets, Phare (multilingual LLM safety benchmark, built with Google DeepMind), RealHarm (problematic agent interactions drawn from real public incidents) and RealPerformance (functional failure patterns), plus patents filed in FR/US/EP.