Measure and limit what AI agents can leak when they reason over confidential data.
About Serinyx
Serinyx builds AI for high-stakes decisions on confidential data. On our decision-intelligence platform, AI agents break a hard question down, research it with checked citations and debate the findings under human review. The result is a verified, cited and signed report.
Our second product gives universities and research institutes an always-current, AI-built view of their research. We were founded in Saarbrücken in 2026 and are now building our engineering team.
The role
You lead our research on AI agents built on large language models (LLMs) that reason over confidential data across organisations. You study what these agents leak and how much, define how to measure it and test defences in experiments with real models. You also help design the secure link between a private deployment on a customer’s own servers and our public cloud platform, which a distributed-systems engineer in the team builds.
Why this role
- An open research problem on the privacy of LLM agents, studied with real models on real tasks.
- Your methods go straight into a working system, built by an engineer in the team.
- You lead the research agenda and write up the results for a peer-reviewed venue.
What you will do in your first year
- Define a measurable bound on what a debate between LLM agents can leak, and build and validate an estimator for it.
- Design, with our CTO, the protocol that agents use to exchange arguments between the private and public sides, and test that protected data cannot cross it directly.
- Build the evaluations and attack tests that show an outside auditor how well our checks protect confidential data.
- Set the measurement method, run the main leakage measurements and write up the results for a peer-reviewed venue.
What you bring
- A PhD (or a submitted thesis) in computer science, machine learning or a related field, or equivalent research experience with publications.
- Research on large language models or AI agents, for example on their evaluation, reasoning, multi-agent systems or security.
- Statistics for measurement: estimators, confidence bounds, variance reduction.
- Publications at strong machine-learning, NLP or security venues.
- Enough Python to build and run your own experiments with LLMs, for example with PyTorch, Hugging Face or model APIs.
Nice to have
- Privacy or security of AI systems (for example data leakage, membership inference or prompt injection); differential privacy or information-flow control; formal methods or protocol verification (for example Tamarin, ProVerif or Lean).
How we work
- One team, one backlog. Everyone builds from day one and works directly with our CTO, Dr Raja Naeem Akram.
- Owners, not ticket-takers. Each engineer owns parts of the codebase and leads the AI coding agents working in them: writing their work orders, reviewing every pull request and deciding what merges.
- A typical week. Much of your time goes into specifying and reviewing work done by AI coding agents; the hardest parts you write yourself.
- AI agents take the routine platform work. Infrastructure, deployment and monitoring are done by AI agents under engineering review, so your time goes into design and the hard problems. You still own what you build once it is live.
- Quality is part of the job. You write the tests for your own work and keep CI green. Done means tests pass, security scans are clean, docs are updated and container images are signed.
- Two-week cycles. Planning on day one, a demo on day ten and a short written retro.
- English first. English is our working language; German is welcome but not required.
What we offer
- Real ownership. An early role in a small team: you shape the product, the code and how we work, not a ticket queue.
- A modern stack. Python, FastAPI, Temporal, Kubernetes, React with TypeScript, and the best AI models for each task.
- 30 days of holidays per year.
- International applicants welcome. If you come from outside the EU, we support you through the work-permit process, usually the EU Blue Card.
Equal opportunity
- We welcome applications from everyone, whatever their ethnic origin, gender, religion or belief, disability, age or sexual identity. You do not need to include a photo or your date of birth.
Research Scientist, Formal Methods & AI Security (all genders) Arbeitgeber: Serinyx
Serinyx bietet ein dynamisches Arbeitsumfeld, das sich auf die Umsetzung von KI-generierten Forschungsergebnissen in umsetzbare Entscheidungen konzentriert. Mitarbeiter haben die Möglichkeit, an innovativen Projekten zu arbeiten, die die Qualität von Prognosen anhand realer Ergebnisse messen.