Giskard

What Is Giskard?

Giskard AI SAS, incorporated in Paris in October 2021 under SIREN 904 773 421, builds open-source tooling for testing and red-teaming large language model agents and retrieval-augmented generation (RAG) systems. Alexandre Combessie and Jean-Marie John-Mathews co-founded the company around a specific and narrow niche. Giskard does not build LLM applications themselves — it builds the scanner that finds what's broken in them before you ship.

That distinction matters more than usual in this category, because it is easy to confuse Giskard with an LLM orchestration framework. It is not one. Giskard's open-source library, released under Apache 2.0 and carrying roughly 5,700 GitHub stars, runs automated vulnerability scans against an existing LLM agent or RAG pipeline — probing for hallucination, prompt injection, sycophancy, bias, and data disclosure. The commercial Giskard Hub adds a managed or self-hosted dashboard, human-in-the-loop review workflows, and a signed-off "go/no-go" deployment recommendation from Giskard's own security team.

Giskard picked up independent visibility in February 2025 when it co-launched Phare, a multilingual LLM safety benchmark, at the Paris AI Action Summit with Google DeepMind as a research partner. Phare scores models on hallucination, bias, harmfulness, and jailbreak vulnerability — giving the company a public research credential that sits alongside its commercial product, rather than relying purely on vendor-reported claims.

The company's backers hint at where it sits in the French AI ecosystem. Alongside lead seed investor Elaia and Bessemer Venture Partners, Giskard's own site lists individual investors including Julien Chaumond, CTO of Hugging Face. Also on the list: Charles Gorintin, co-founder and CTO of Alan, and a co-founding advisor at Mistral AI. That network of AI-industry insiders, rather than a large late-stage raise, is the clearest signal of how Giskard has built credibility so far.

Key Features

Automated Vulnerability Scanning

Point Giskard's scan function at a model or agent and it runs a battery of tests drawn from internal knowledge bases, published security-vulnerability taxonomies, external resources, and prompt-template libraries. Detected issues span both security failures — prompt injection, sycophancy attacks, unintended data disclosure — and quality failures such as hallucinations, internal contradictions, and inappropriate refusal behaviour. Because the scan works in black-box mode against an API endpoint, it can test an agent without needing access to model weights or internals.

RAGET: RAG-Specific Evaluation

RAGET (RAG Evaluation Toolkit) is Giskard's answer to a problem most RAG teams solve manually: generating a realistic test set. Feed it a knowledge base, and RAGET generates questions, reference answers, and reference contexts using a chain of LLM calls, then evaluates an agent's real answers against that test set. Critically, RAGET maps failures back to a specific RAG component — retriever, generator, or knowledge-base chunking — rather than just reporting a pass/fail score. That component-level detail is what makes debugging a broken RAG pipeline tractable instead of a guessing game.

Autonomous Red-Teaming

Giskard's red-teaming probes run without hand-written test cases, generating adversarial inputs designed to trigger unsafe or incorrect behaviour and then classifying the response against a severity-ranked taxonomy. This is the feature set that most directly overlaps with dedicated AI red-teaming vendors. It is also where Giskard's open-source distribution gives it an edge, since teams can run these probes continuously in CI/CD rather than as a one-off pre-launch audit.

Giskard Hub

The commercial Hub wraps the open-source scanning engine in a dashboard built for teams, not just individual engineers. It ranks vulnerabilities by severity, generates remediation guidance for each finding, and supports continuous post-deployment monitoring rather than a single point-in-time report. Hub deployments can run fully on-premises for organisations that cannot send prompts or outputs to a third-party service. Giskard states the Hub also supports SOC 2 Type II-aligned controls, role-based access control, and audit trails, with a choice between EU and US data residency.

The Phare Benchmark

Phare is a separate but related project: an independent, multilingual benchmark that evaluates any LLM — not just ones tested through Giskard's product — across hallucination, bias, harmfulness, and jailbreak resistance. Google DeepMind's involvement as a research partner gives Phare a credibility signal that a company of under 20 people could not generate alone. It also doubles as a marketing asset, demonstrating Giskard's evaluation methodology in public.

Pricing

Giskard's open-source library is free, full-featured, and requires no Giskard account — you can run scans and RAGET evaluations entirely offline against your own infrastructure. This mirrors the model used by fellow French AI-tooling company deepset with Haystack: the free tier is a genuine product, not a crippled trial designed to push you toward the paid layer.

Giskard Hub, the commercial dashboard-and-monitoring product, publishes no price list. Every quote requires a sales conversation, and pricing presumably scales with usage, seats, or deployment model, though none of that is disclosed publicly. For a company this size, an opaque enterprise price list is a normal but real friction point — buyers comparing options can't shortlist on cost the way they can with, say, Langfuse's published tiers.

EU Compliance & Privacy

Giskard AI SAS is a French company under full EU jurisdiction, which puts GDPR enforcement and data-subject rights squarely within the French data protection authority's remit rather than a foreign regulator's. Hub customers can choose between EU and US data-residency regions. The open-source library, meanwhile, can run entirely self-hosted, so prompts and outputs never need to leave your own infrastructure if you don't want them to.

Giskard's own website states that the Hub maintains SOC 2 Type II-aligned controls and HIPAA-oriented safeguards. As with any vendor-stated compliance claim, enterprise buyers in regulated sectors should request the underlying audit documentation rather than taking the marketing page at face value.

Who It's Best For

Teams already running LangChain, Haystack, or a custom LLM stack who need a testing layer are the clearest fit. Giskard sits alongside your existing framework rather than replacing it, scanning whatever agent or RAG pipeline you've already built.

RAG teams debugging inconsistent answer quality get direct value from RAGET's ability to isolate whether a failure sits in retrieval, generation, or knowledge-base chunking, rather than treating the whole pipeline as one black box.

Security and red-team functions inside AI-heavy organisations benefit from the autonomous probes and severity-ranked vulnerability reports, particularly where continuous CI/CD-integrated testing matters more than a one-time audit.

Enterprises needing dedicated LLMOps or agent-building infrastructure should look elsewhere first. Giskard does not replace LangChain or LlamaIndex as a framework — teams building the application layer need one of those, or deepset's Haystack, alongside Giskard, not instead of it.

The Verdict

The gap it fills is real and underserved: a European, open-source testing and red-teaming layer for LLM agents and RAG systems. Independent credibility through the Phare benchmark and Google DeepMind partnership backs that up. It is honest about what it is not — an LLM framework, an orchestration layer, or a LlamaIndex-style pipeline builder — and buyers who understand that boundary get real value from it. The risks are the ones that come with any company of under 20 people: support depth, roadmap continuity, and an unpriced enterprise tier that requires a sales call before you know what it costs.

Frequently Asked Questions

Is Giskard open source?

Yes. Giskard's core scanning and testing library is open source under the Apache 2.0 licence, hosted at github.com/Giskard-AI/giskard with roughly 5,700 stars. The commercial Giskard Hub adds managed dashboards, human review workflows, and SLA-backed support on top of the open-source engine.

Does Giskard replace LangChain or LlamaIndex?

No. Giskard is a testing and red-teaming layer you put around an existing LLM stack, not an application or orchestration framework. It is a genuine European alternative to LangChain's commercial LangSmith evaluation tooling specifically, not to the LangChain framework itself. Teams looking for a European framework peer to LangChain or LlamaIndex should look at deepset's Haystack, and teams wanting an EU observability layer should look at Langfuse.

What is RAGET?

RAGET (RAG Evaluation Toolkit) is Giskard's module for testing retrieval-augmented generation systems. It automatically generates realistic questions, reference answers, and reference contexts from a knowledge base, then evaluates an agent's answers against them to isolate whether failures come from retrieval, generation, or knowledge-base quality.

What is the Phare benchmark?

Phare is an independent, multilingual LLM evaluation benchmark that Giskard launched at the Paris AI Action Summit in February 2025, with Google DeepMind as a research partner. It scores models across hallucination, bias, harmfulness, and vulnerability to jailbreaking, giving buyers a third-party reference point beyond vendor-reported benchmarks.

Is Giskard GDPR compliant?

Yes. Giskard AI SAS is a French company (SIREN 904 773 421) incorporated in Paris in October 2021 and is fully subject to EU data protection law. The Giskard Hub offers a choice of EU or US data residency, and the open-source library can be self-hosted entirely within your own infrastructure for maximum control.