Edgeless Systems

For those of you who regularly work with sensitive data in Excel, Word, or PowerPoint: we have some exciting product news to share! Today, we are launching our Desktop Proxy, which can turn your Claude plugins within the Microsoft ecosystem into confidential plugins. The solution keeps the UI as is, but routes the underlying model calls to a secure Privatemode endpoint instead (including remote attestation). Simply download it, set it up (it takes just 5 minutes, see https://lnkd.in/eFprtMBt for instructions), and keep the proxy running while you work in Office. You can find the download and more information here: https://lnkd.in/eEMmbCcG

Together with TWT Group we are hosting a webinar on 03.11 on Confidential AI (in German). The starting point will be familiar to many teams: the prototype works, but it gets stuck on one question. May it see real data? The answer depends on the whole chain, not only on the model provider. Rüdiger Pläster (Chief AI Officer, TWT) and Thomas Strottner (Edgeless Systems) take one assistant for internal company knowledge and walk through it live, using fictional documents: ✅ What confidential computing protects, and where it stops ✅ How Privatemode is integrated for protected processing ✅ What still needs to be solved before production: data sources, access rights, logging, answer quality 60 minutes, free of charge, recording included. 📅 03.11.2026, 11:00 to 12:00 CET Registration: https://events.twt.de/

New case study alert: See how DIGIBIZ now reads and tags a 320-page paper file in under 20 minutes, for €0.39 in model costs. S + P DIGIBIZ GmbH builds software for the SCHNEIDER + PARTNER consulting group, where auditors, tax advisors, and lawyers work under professional secrecy. Until this spring, AI tools were deliberately off-limits for client files. The legal team's position was clear: a Data Processing Agreement (DPA) alone is not enough. It defines what a provider may do with the data, but with common AI services, prompts are still processed in plaintext on the provider's servers. In our recent case study, we share how Privatemode AI is applied to process documents inside a hardware-isolated environment whose memory stays encrypted. A proxy on the client side verifies the environment through remote attestation before any data is sent. You can find the case study (in German) here: https://lnkd.in/etsWsFus

Big announcement: we are launching Privatmode Decisions - a confidential alternative to Jev. With our new reduced pricing on GLM-5.3 Flash, we can provide blazing fast inference (avg. 156ms per decision in Europe) for a cost of 0.06€ per 1k decisions, on par quality with Jev on a benchmark of 28 datasets. The benchmark is published open-source on GitHub, a playground can be found in our blog post: https://lnkd.in/eXCGFaB9 Grab an API key and try it out yourself!

Have you already tried our recent additions of GLM-5.3 and GLM-5.3 Flash to Privatemode? GLM-5.3 is the model with the fastest growing demand we've seen on our platform so far. It matches frontier LLM capabilities. The smaller "Flash"-sibling comes with vision capabilities, so you can also use it for tasks like document understanding and more. You can find more details on our models page: https://lnkd.in/e7XSRPMV

Next week, Ömer Tekin and Johannes Hötter of Edgeless Systems are speaking at the Hamburg Insurance Innovation day (15.09).   Insurance companies need to follow strict regulations and process highly sensitive data. The use of AI so far is often limited to publicly available data, and many insurers are looking into options around expensive local GPU infrastructure.    In their talk, Ömer and Johannes are presenting confidential AI as an option for insurers to accelerate their AI journey. If you are in town and working in insurance, make sure to not miss this event! Link: https://lnkd.in/eAXwV8SQ

If you're programming with JS / TS and want to process sensitive data with AI, we've got some exciting updates for you! We just released a new SDK, which allows to securely verify and connect to Privatemode from a JavaScript / TypeScript client application with an OpenAI-compatible client. You can find it in our changelog (docs): https://lnkd.in/eBDVJ87W Or directly on npm: https://lnkd.in/eSjsX8EF

Wir sehen täglich die sensibelsten Daten unserer Kunden. Trotzdem nutzen wir AI. Nicht über Vertrauen, sondern über Kryptografie. Übliche Cloud-AI-Dienste versprechen Datenschutz per AGB. Ob Prompts geloggt oder fürs Training genutzt werden, kann von außen niemand prüfen. Deshalb nutzen wir Privatemode von Edgeless Systems. Prompts werden clientseitig mit AES-256 verschlüsselt und erst in einer Confidential-Computing-Umgebung entschlüsselt, abgeschirmt durch AMD SEV-SNP und NVIDIA H100. Im Hauptspeicher liegen sie nie im Klartext. Remote Attestation beweist vor dem ersten Prompt kryptografisch, welcher Code dort läuft. Weder der Cloud-Provider noch Edgeless Systems selbst kann unsere Daten lesen. Datenschutz per Vertrag ist ein Versprechen. Datenschutz per Hardware ist ein Beweis. https://lnkd.in/eD_sa2Ck #ConfidentialComputing #Datenschutz #KI #CyberSecurity #Pentest

We are excited to announce our partnership with GitLab to bring frontier AI models via Privatemode AI directly to GitLab customers that self-host their stack. This lets you run state-of-the-art coding agents based on models like Kimi or GLM. No on-prem GPUs required. Your data stays encrypted end-to-end, even during processing. This matters most in highly regulated industries, where frontier models usually fail the security review before they ever reach a developer. Find out more about this partnership on GitLab's Blog: https://lnkd.in/gCz9XP4F Also, feel free to check out our integrations page: https://lnkd.in/gKpDWynn

Confidential containers in Kubernetes: what a real operator exclusion requires At the 21st German IT Security Congress (organized by the Federal Office for Information Security (BSI)), Felix Schuster and Moritz Eckert analyze under which conditions a technical operator exclusion via confidential computing in Kubernetes is achievable. The supporting blog post, including a link to the full article, is available here: https://lnkd.in/dBDqGp3z