A European AI research lab building sovereign-by-default AI systems, and studying how they break.
euraika-labs.net — the editorial surface where we write about what we build. euraika.net — the parent organisation.
Euraika-Labs builds AI infrastructure for organisations whose data, compliance posture, or jurisdiction makes US-cloud LLM stacks a non-starter. The systems run end-to-end inside the customer's own infrastructure under EU jurisdiction. This GitHub organisation hosts the open-source side of that work.
Systems that handle regulated data are only as trustworthy as their failure modes, so part of our work is adversarial: we study how LLM systems fail and publish the controls that survived evaluation.
- Prompt injection — a vector taxonomy and the few mitigations worth the engineering. Read
- Agent containment — capability-scoped runtimes and approval gates for agents that can touch a filesystem, a network, or a keyboard. Paper · How Pan gates tool calls
- Isolation and data flow — gateways that refuse unclassified prompts, and prompt caches that cannot leak across tenants. Gateway · Cache
- Evidence integrity — tamper-evident yet redactable audit trails, and benchmarks that cannot be quietly re-tuned after the fact. Hash chains
| Project | Status | Stack | What it does |
|---|---|---|---|
| pan-agent | stable | Go + Tauri + React | AI desktop agent. Full PC control behind an explicit approval system. Single static binary, no telemetry. |
| pan-ui | stable | Next.js + TypeScript | Self-hosted web workspace for chat, skills, extensions, memory, profiles, and runtime controls. |
| pan-desktop | stable | Tauri + Rust | Native Windows desktop app for the Pan workspace. Hard fork of fathah/hermes-desktop (MIT). |
| personal-assistant | beta | JavaScript | Safety-reviewed Agent Skill for Claude Code and Codex. Intent gate, explicit authorisation for consequential actions, no telemetry. |
| personal-assistant-memory-routing | beta | JavaScript | Privacy-first memory recall across providers. Read-only, bounded, and marked as untrusted data. No automatic writes. |
| Project | Status | Stack | What it does |
|---|---|---|---|
| open-grc-benchmark | alpha | Python | Synthetic regression fixtures for source-grounded GRC assistants. Offline, frozen and hashed, no combined score. |
| ai-slopcheck | stable | Python | Deterministic scanner for AI-style code failures, including error-swallowing and leftover mocks. 72 rules. pip install ai-slopcheck. |
| benchforge | alpha | Rust | macOS workbench for reproducible benchmarks of local and cloud LLMs, with inspectable artifacts. |
Open research projects. No releases or results yet; each README states the question, threat model, method and what it will not claim. Contributions and discussion welcome.
| Project | What it studies |
|---|---|
| skillcheck | Deterministic scanning of Agent Skills, hooks and MCP manifests for exfiltration, hidden instructions and over-broad capability. |
| injection-bench | A reproducible prompt-injection corpus for tool-using agents, organised by vector, with mechanical success checks. |
| memory-poisoning | How injected instructions persist through agent memory systems, and which defences hold. |
| approval-gate-bypass | Adversarial evaluation of desktop-agent approval gates, starting with our own. |
The commercial products ship under commercial licences but are designed to run end-to-end inside the customer's own infrastructure:
- Aegis — AI-powered compliance command center. Single-tenant SaaS for SOC 2, ISO 27001, NIS2, and EU AI Act compliance.
- Hermes (gateway) — Production LLM gateway. Multi-runtime, with a model catalogue, governance hooks into Aegis, and routing that respects EU jurisdiction by default.
- Vane (Camiel) — Privacy-focused AI answering engine with verified synthesis and citations.
More on each at euraika-labs.net/about.
A workload is sovereign under our definition when an organisation can run it end-to-end without any of its data, prompts, model weights, vector indices, or operational metadata leaving its jurisdiction, its network, or its legal control — and when this is the default configuration, not an enterprise tier. Full definition.
Found a security issue in one of our repositories? Please report it privately — see SECURITY.md. Do not open a public issue.
- Email: hello@euraika.net
- Website: euraika-labs.net
- Parent organisation: euraika.net
The team: Bert Colemont, Andrei Lavrenov, Laurence Bedoret, Corey Stables, Amin Zakkane.