Giskard vs Patronus AI

Giskard is a Paris-based company founded in 2021 by Alex Combessie, Jean-Marie John-Mathews and Andrei Avtomonov. Combessie and John-Mathews are now co-CEOs, and Giskard says its co-founders previously built AI systems at Dataiku and Thales. It sells software and services for testing and securing LLM agents. Its first funding round, €1.5 million led by Elaia with Bessemer Venture Partners, was announced in December 2022. Giskard says it also received a €3 million strategic investment through the European Commission's EIC Accelerator in 2023 and a €3 million Bpifrance grant in 2024 for LLM evaluation research with Mistral AI, Artefact, INA and BnF. CORDIS lists a €2.5 million EU contribution to its EIC project, which ran from September 2023 to August 2025. Giskard Hub is the paid platform. Continuous red teaming runs multi-turn attacks that adapt to the agent's answers, generated from the company's own documents and knowledge bases. It tests for hallucinations, stereotypes and discrimination, harmful content, disclosure of personal information and prompt injection. Testing is black-box: the agent only needs an API endpoint, and coverage is limited to text-to-text conversational agents. The Hub also schedules agent evaluations and manages test datasets, custom checks and human review, with alerts, SSO, role-based access and audit logs. It runs as SaaS, in a private cloud or on premises, with EU or US data processing. Giskard Guards, launched in April 2026, is a runtime guardrail layer that runs in the customer's infrastructure. It inspects prompts, responses, tool calls and multi-step reasoning, and its policies are written as code in OPA/Rego. Giskard also sells AI security audits that end in a go/no-go report, and it is offering a free trial of a new Agentic Red Teamer to a limited number of users. The open-source library is Apache 2.0 licensed. Version 3, a rewrite for multi-turn agent testing, became stable in August 2026. Version 2 is no longer actively maintained, and tabular models have no v3 equivalent. Giskard also publishes Phare, an LLM safety benchmark launched in February 2025 with Google DeepMind as research partner. Logos on its site include BNP Paribas, AXA, BPCE, Société Générale, Crédit Agricole, Michelin, Decathlon, L'Oréal and Doctolib. **AI compliance coverage** Guards includes an EU AI Act policy pack for prohibited practices such as social scoring and real-time biometric identification, and Giskard says Guards covers OWASP LLM risks. It doesn't map its products to the NIST AI RMF or ISO 42001. Its trust center, run on Bastion, lists GDPR and a SOC 2 Type 2 report and shows HIPAA as in progress. **Who it's for** Enterprises, especially in European banking and insurance, that run customer-facing LLM agents and want testing and runtime controls they can deploy on premises.

Patronus AI was founded in 2023 by Anand Kannappan (CEO) and Rebecca Qian (CTO), who both previously worked on AI at Meta, and is based in San Francisco. It launched in September 2023 as an LLM evaluation company for regulated industries. It raised a $3 million seed round led by Lightspeed Venture Partners, then a $17 million Series A led by Notable Capital in May 2024. In June 2026 it raised a $50 million Series B led by Greenfield Partners, with Lightspeed, Notable Capital, Datadog and Samsung, bringing its total to $70 million. **What it sells now** Its homepage now says: "Patronus AI is a research lab developing training data and simulated environments for frontier AI." The products page offers Frontier Training Data: tasks with reference solutions and graders, and ready-built or custom simulated environments for software development, computer use and knowledge work, delivered to run in a lab's own training infrastructure. In December 2025 it launched Generative Simulators and RL Environments, which generate tasks and rules and adapt them as an agent learns. In June 2026 it announced Digital World Models, replicas of websites and company systems where agents can practice. Patronus says most leading frontier labs and hyperscalers are customers but names none, and it reports revenue growth of more than 15 times over the past year. **The evaluation platform** Patronus's documentation still describes the evaluation product the company started with. It includes an API and SDK with off-the-shelf evaluators (among them Lynx for hallucination detection and the small Glider evaluator model), custom evaluators, experiments, logs, tracing and Percival, launched in May 2025 to find failures in agentic systems and suggest fixes. Evaluation Runs and the Experimentation SDK are Enterprise features. An older landing page names OpenAI, HP and Pearson as customers and offers on-premise deployment. The homepage, products page and pricing page no longer mention the platform. **AI compliance coverage** Patronus doesn't map its products to any AI regulation or standard and publishes no security certifications. Its older landing page says it is vetted by third-party security firms each year. **Who it's for** Frontier AI labs and large AI teams that buy training data and agent environments. Teams looking for an evaluation or observability tool to support compliance work should confirm that the evaluation platform is still sold and supported before buying.