Giskard and the EIC Accelerator: Stress-Testing AI Agents

Giskard, registered in Paris in October 2021, sells software that automatically tests AI models and agents for errors, bias and security holes such as prompt injection, alongside an open-source Python library. Its founders came from Dataiku and Thales. It was selected in the EIC Accelerator's March 2023 cut-off and received an EU grant of €2.50 million for "Quality Assurance for AI", a €3.73 million project that ran from September 2023 to August 2025, according to CORDIS. The objective was to validate its testing tool, extend it to time series and computer vision, and become "the leading European software provider" for the EU AI Act.

The final report says Giskard "successfully completed its Quality and Security testing platform", secured two patents, and built open-source scans for LLM, text, tabular and vision models, an enterprise platform for testing generative AI agents and a platform to assess conformity with the AI Act, according to the CORDIS reporting page. In May 2026 Giskard launched Giskard Guards, runtime guardrails for AI agents, and said its testing platform had been used by Mistral, Google DeepMind, BNP Paribas, AXA and Doctolib, according to Tech.eu.

Where Giskard Stands Now

Giskard is active in the French company register, which puts its workforce at 10 to 19 employees, according to recherche-entreprises. The free register data holds no usable revenue or profit figures. Official BODACC notices record increases in its share capital in January and September 2026, with no amounts in the notices. The open-source library had 5,847 stars on GitHub, with commits in the last week of September 2026, and 12,840 downloads in the preceding 30 days on PyPI. A paper by its research team on stereotypes in frontier LLMs was accepted at NeurIPS 2026, according to the company website.

The Technology: Attacking AI Before Others Do

Giskard's tools generate adversarial tests, such as attempts to make a banking chatbot leave its role or leak instructions, and check the answers for hallucinations, bias and unsafe actions. The newer Guards product sits in front of a live agent and inspects its tool calls and reasoning steps. John-Mathews told Tech.eu that generic guardrails can block up to 40% of legitimate requests because they lack business context. That figure is his, not an independent measurement.

What the EIC Award Promised, and What the Record Shows

The award promised a validated AI testing platform extended to more model types, and an AI Act compliance offer. The final report says all of this was built, the open-source library is public and used, and enterprise users are named. The report does not give revenue or customer numbers. Giskard's own website describes the EIC money as a "strategic investment of 3 M€ from the European Commission"; CORDIS records a €2.50 million grant under the grant-first route.

1. Testing platformDeliveredCompleted, per report.
2. More model typesDeliveredLLM, tabular, vision scans.
3. AI Act compliance toolDeliveredBuilt, per report.
4. Commercial scaleUnverifiedRevenue not public.

The Money

The EU paid 67% of the project. Giskard's only announced equity round was €1.5 million in December 2022, led by Elaia with Bessemer Venture Partners. In 2024 it won a share of a €3 million Bpifrance grant for LLM evaluation research with Artefact, INA, BnF and Mistral AI, according to the company's About page. On disclosed figures, public money exceeds private equity. The 2026 capital increases have not been announced.

What Cannot Be Checked

Why Europe Should Care

The AI Act requires testing and risk management for high-risk AI systems, and European companies deploying chatbots need to know how they fail. Giskard is one of few European vendors in AI red-teaming, a field with many US competitors. On disclosed figures, it has built its product and named user base on mostly public money; whether it grows into a sizeable company depends on private funding it has not disclosed.

The Verdict

Giskard and the EIC have not been asked for comment for this article. It is based entirely on public records and published statements, linked throughout.