

Foresight’s AI Salon | Engineer AI safety workshop
When an AI agent acts for your team, what actually stops it from doing the wrong thing, including when the model itself is the problem? This summer's OpenAI–Hugging Face incident showed what happens when containment depends on a single filtered network path and nobody is watching. This salon is about the engineering side of the answer: safety as a system property you build, test and verify.
Join Foresight’s AI Salon with Wirken (gebruder/wirken), an open-source gateway written in Rust that sits between your team’s messaging channels and the AI agents acting for them.
Every action an agent proposes passes a deterministic permission gate. Senders and approvers are allowlisted, each credential is bound to the destinations it may reach, and every session is recorded in a signed audit chain you can verify offline. Its controls map to the OWASP Top 10 for Agentic Applications.
The model proposes and the gateway decides, so a small open model running on your own laptop can be trusted to act. We'll also look honestly at what the gateway doesn't cover: model intent, side channels, poisoned inputs, approver fatigue and real-time detection.
This salon is for you if:
you work on AI safety, security or control, or are moving into it, and take risks from increasingly autonomous agents seriously
you build, deploy or secure agent systems and want controls you can verify, not just promises
you want to keep going afterwards: piloting agent controls in your team or research, contributing, or presenting at a future salon
you are part of a small organization or community group that wants to run its own agent on its own machine, even without a security team
This is a working session, not a general AI networking evening. Places are limited and we review registrations.
Before you come (required for the breakouts): a laptop with Docker running, Wirken installed (see the repo README) and Ollama with a small open model. A hosted model API key also works, but the session is built around local.
You'll leave with: a running Wirken instance with your own permission policy, your agent use case mapped against the OWASP Agentic Top 10, and a place in the Wirken community group in Berlin: people piloting it in their own team, contributing, and running the next session.
Wednesday, November 25th · 18:00–21:00 CET
CIC Berlin
17:45-18:00 doors open (be punctual, main doors are closed afterwards)
18:00–18:20 drinks and meet other researchers and builders
If everyone brings some food, we can share together.
Plates and cutlery will be provided.
18:20–18:40 Why architectural safety beats safety pledges (Davi Ottenheimer)
18:40–19:00 Q&A
19:00–20:30 Breakouts. Pick one. All three run on a small local model.
Write the rules. Build a permission policy for your own agent use case, then watch the gate allow, ask and refuse.
Read the record. Reconstruct what an agent did from the signed audit chain alone, offline, and verify nobody altered it.
Break it. Red team the gate. What gets past it, and what does the record show afterward?
20:30–20:45 Each group reports what it found.
20:45–21:00 Social time and cleaning up together
Registration: request to join and answer four short questions. We'll confirm places by 23 November.
About the speaker:
Davi Ottenheimer has led security engineering for three decades, at Yahoo!, VMware, MongoDB and EMC, and earlier at UCSC and UARC-NASA. He worked for six years with Tim Berners-Lee on his vision of the future web. He is co-author of Securing the Virtual Environment, on the IANS faculty, and Managing Director of Ottenheimer GmbH in Berlin, where he builds Wirken.
He argues that the "saccharin safety pledges" of AI labs are marketing, and that the real question is architectural: can you stop an agent, and can you prove afterward what it did? He has written with Bruce Schneier on robot safety and on data integrity, and publicly challenges AI labs when their safety claims and their deployments diverge. His thesis: safety is not a prompt. It is a system property you engineer and verify.
Salon discussions are held under Chatham House Rule (don’t connect people to ideas when discussing them outside of this workshop).
Foresight’s AI Nodes offer grant funding, local compute and community hubs for AI for Science and Safety projects in Berlin and the Bay Area.