European AI Security Platforms

#1 Mindgard

London, United Kingdom Founded 2022 Pricing on request (demo-led) Demo on request

Best for: Security teams that want proof of what an attacker could do to a model or agent before it ships

• Operating company. Mindgard Ltd
• Jurisdiction. United Kingdom (adequacy decision, outside the EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer AI attack simulation, manual AI penetration testing

Mindgard sells AI security as a sequence of attacks. Its platform runs Recon to map what a system can reach, Attack to find exploitable weaknesses, Remediate to act on the gaps and Assure to rerun the tests whenever the system changes, so one run can double as a release gate and as an audit trail. Entry points are an API, CI/CD pipelines, Burp Suite or a one-click workflow.

Mindgard Ltd is a London company that says it was founded in 2022 on research by Dr Peter Garraghan at Lancaster University. Its terms are governed by the laws of England, Northern Ireland, Scotland and Wales. Hosting region, self-hosting and certifications were not stated on the pages reviewed, and prices are given only after a demo, so ask for both in writing.

Standout feature. Mindgard publishes more than 150 AI vulnerability disclosures, so a buyer can judge its research before booking a call.

#2 NeuralTrust

Barcelona, Spain Free: 100,000 requests/month. Standard from $49/month (listed in USD) plus usage. Enterprise on request. Demo on request

Best for: Teams that want an agent gateway and runtime guard they can start using on a free plan

• Operating company. NeuralTrust, S.L.
• Jurisdiction. EU (Spain)
• Where the data sits. Free plan: US or EU data residency; private deployment on the enterprise plan
• Source code. Closed source
• Replaces. HiddenLayer AI guardrails and runtime security

NeuralTrust divides its offer into TrustGate and TrustGuard. TrustGate routes traffic from agents to models and tools and applies model routing, fallbacks, rate limits, tool permissions and budget caps; TrustGuard inspects interactions for prompt attacks and data leakage. Because both sit in the request path, one policy can cover homegrown agents, third-party agents and endpoint use.

NeuralTrust, S.L. is the contracting company named in the terms of service, with a Barcelona address in the privacy policy, which puts it under Spanish and EU law. Its founding year and registered office were not confirmed from its pages. The latency figure of under 100 ms and the 99 percent detection accuracy on the site are the vendor's own claims, and prices are quoted in US dollars.

Standout feature. NeuralTrust is the only vendor in this ranking that prints a free allowance and a price per request on its pricing page.

#3 Advai

London, United Kingdom Pricing on request (call booking) Demo on request

Best for: Regulated firms that must give an approver written evidence before an AI system goes live

• Operating company. Advai Limited
• Jurisdiction. United Kingdom (adequacy decision, outside the EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer attack simulation and model monitoring

Advai describes its service in three verbs: choose a model for cost and latency, test it for manipulation, data handling failures and reliability, and monitor it in production for drift, policy breaches and cost spikes. Testing ends with go-live thresholds and an evidence pack aimed at the person who signs off, and coverage extends to agents, language models, computer vision, time series and multimodal systems.

Advai Limited is registered in England and Wales under number 12540124 with a registered office in London, and its platform page states ISO 27001 and Cyber Essentials Plus certification. The sectors it names are financial services, insurance, legal and defence. It says testing is its only business, which it offers as the basis of its independence, and a managed onboarding sets up the first tests. Hosting region and prices are not published.

Standout feature. Advai turns test results into an evidence pack written for the person who has to approve deployment.

#4 Giskard

France Founded 2021 Open source free; Professional on request Demo on request

Best for: Developer teams that want to test models and agents before release, starting from open source

• Operating company. GISKARD AI SAS
• Jurisdiction. EU (France)
• Where the data sits. On-premise, private cloud or SaaS with data residency options
• Source code. Closed source
• Replaces. Weights & Biases Weave evaluations, Arize, Deepchecks

Giskard generates adversarial and quality tests for machine learning models and agents. The security side covers sycophancy, prompt injection, data disclosure and inappropriate content, while the quality side covers hallucinations, contradictions, omissions and inappropriate denials. The Professional tier lists more than 50 automated adversarial probes, including multi-turn attacks, and coverage of agent risk frameworks such as OWASP.

GISKARD AI SAS builds in France; its About page says it was founded in 2021 by people who had worked at Dataiku and Thales. The library sits on GitHub under Apache-2.0, though its README says v3 is a fresh rewrite, and the enterprise tier lists on-premise, private cloud or SaaS with data residency, SSO, role-based access and SOC2, HIPAA and GDPR compliance. It does not track training runs or serve models.

Standout feature. Giskard is the only tool here whose test library is open source, so a developer can run it locally without a sales call.

#5 QuantPi

Saarbrücken, Germany Founded 2020 Pricing on request (demo-led) Demo on request

Best for: Enterprises that need a pass or fail result for every release of a model, RAG pipeline or agent

• Operating company. QuantPi GmbH
• Jurisdiction. EU (Germany)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer attack simulation, ad hoc model test scripts

QuantPi calls its product a validation layer and puts statistics at the centre. Each test result carries a confidence interval so a reviewer sees how far to trust the number, and the site describes a method for using an LLM as a judge while measuring how reliable that judge is. One engine is meant to cover foundation models, RAG pipelines, tabular models, object detection and agentic workflows.

QuantPi GmbH has an imprint at Halbergstr. 4 in Saarbrücken, and its company page says it was founded in 2020 as a spin-off from the CISPA Helmholtz Center for Information Security and Saarland University. Validation can be called through MCP, GitHub Actions or Jira, and a second workflow maps a regulatory standard to a test suite and pushes the evidence to a GRC tool. Prices, hosting and self-hosting were not published.

Standout feature. QuantPi reports how reliable each test result is, down to the judge model used to grade it.

#6 Adversa AI

Tel Aviv, Israel Pricing on request (demo) Demo on request

Best for: Security teams that govern coding agents or want an outside red team for LLM applications

• Operating company. Adversa AI Ltd
• Jurisdiction. Israel (outside the EU/EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer AI red teaming

Adversa AI offers two things: continuous red teaming of LLM, agentic, generative and MCP systems, and a coding agent security platform that combines a gateway with a lightweight sensor on developer machines. The platform records what a developer agent reads and what it then runs, judges the whole chain rather than single steps, and blocks dangerous chains before they finish, across several agent brands.

Adversa AI Ltd gives an address on Rothschild Boulevard in Tel Aviv in its privacy policy, and the site publishes a data processing agreement and a security and trust centre. On 4 June 2026 the company released AIRQ, a rating of 100 enterprise AI agents with an open source methodology. Funding and awards on the site are vendor statements, and incorporation details, prices and hosting are not given.

Standout feature. Adversa AI scores an agent session as one chain of actions, so a single odd step is context rather than an alert.

#7 Lakera (Check Point)

Tel Aviv (Check Point), Israel See website (pricing page not readable without scripts) Demo on request

Best for: Developers who want a guardrail API with a choice of processing region or a container of their own

• Operating company. Lakera (Check Point Software Technologies Ltd)
• Jurisdiction. Israel (outside the EU/EEA)
• Where the data sits. Processing endpoints in the EU, USA and Singapore; storage region configurable
• Source code. Closed source
• Replaces. HiddenLayer AI guardrails

Lakera Guard screens what goes into and comes out of an LLM and returns a verdict. The documentation lists prompt defense, content moderation, data leakage prevention, malicious link detection and agent behavior defense, with custom guardrails, allow and deny lists and per-project policies, and the same platform adds red teaming through a Management API and SDK plus agent discovery.

The documentation now brands the product Check Point AI Guardrails, and the terms linked from lakera.ai name Check Point Software Technologies Ltd, registered in Tel Aviv, as the company behind it. The EU endpoint processes in eu-south-2 and storage is configured apart from processing, but the docs state that requests are processed wherever they are sent. The pricing page needs scripts to render, so no price is given here.

Standout feature. Lakera lets you choose where a prompt is processed and where its log is stored as two separate settings.

#8 LatticeFlow AI

Zurich, Switzerland Pricing on request Demo on request

Best for: AI risk teams that must tie technical test evidence to named frameworks

• Operating company. LatticeFlow UK Ltd (LatticeFlow AI)
• Jurisdiction. Switzerland (adequacy decision, outside the EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer attack simulation, manual AI governance checklists

LatticeFlow AI is organised in four modules: Atlas holds a public registry of 40+ frameworks, 400+ controls and 100+ risks; Evaluate runs more than 100 ready-to-run evaluations; Secure runs an adaptive red teaming agent against data leakage, unauthorised actions, privilege abuse, goal hijacking and denial-of-wallet, aligned with OWASP and MITRE ATLAS; Govern tracks risks and actions. Evaluations are written as YAML specs so the method can be audited.

The About page says the founders are professors and researchers from ETH Zurich, and the footer lists offices in Zurich, San Francisco, Dublin and Sofia. The privacy policy names LatticeFlow UK Ltd, so the entity that signs a customer contract needs confirming. Pricing is on request and hosting was not stated. Because security is one module, the product overlaps with Giskard on agent testing.

Standout feature. LatticeFlow AI begins with a framework control and works down to the evaluation that tests it.

#9 Safe Intelligence

London, United Kingdom Pricing on request (demo) Demo on request

Best for: Teams that run vision or classic ML models in safety-critical settings and also deploy agents

• Operating company. Safe Intelligence Limited
• Jurisdiction. United Kingdom (adequacy decision, outside the EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer model attack simulation

Safe Intelligence works at the model level. Its Validate step finds fragilities and counterexamples by analysing performance over whole regions of input space and formally verifying regions against broad classes of perturbation, for deep networks, decision trees and random forests. A repair step then reduces the fragilities and lowers variance in trained models, and Monitor watches for new ones over time.

Safe Intelligence Limited is registered in the United Kingdom according to its terms, and the privacy notice gives a London address. Spec27.ai, introduced on 8 April 2026, adds spec-driven validation for AI applications and agents, with test sets generated for deep coverage and tracking of agents across an organisation. Prices, hosting and certifications are not on the pages reviewed.

Standout feature. Safe Intelligence is the only vendor here that repairs a model after its verification finds a weakness.

#10 Alice (ActiveFence)

Israel, Israel Pricing on request Demo on request

Best for: Companies whose main AI risk is harmful or policy-breaking output

• Operating company. ActiveFence Ltd (d/b/a Alice)
• Jurisdiction. Israel (outside the EU/EEA)
• Where the data sits. Not stated on the pages reviewed
• Source code. Closed source
• Replaces. HiddenLayer AI guardrails and red teaming

Alice approaches AI security from content safety. Its WonderSuite platform offers red teaming, red teaming in production and custom guardrails, and the site argues that off-the-shelf guardrails do not know a customer's policies, so it trains them on the customer's own rules. The offer is grouped by audience: models, apps and agents, online platforms, and personal AI.

Alice is the trading name of ActiveFence Ltd; the terms are governed by Israeli law with courts in Tel Aviv-Jaffa, and no postal address was found on the policy pages. The company describes its Rabbit Hole engine as a network of in-house researchers and a database of billions of examples, which is a vendor description. Solution pages cover the EU AI Act, NIST and OWASP. The brand is recent, so older contracts may say ActiveFence.

Standout feature. Alice builds its guardrails from the customer's written policies instead of shipping a generic rule set.