Our priorities
Alignment and agent safety
How to keep the goals and behaviour of AI systems aligned with human intentions, also when they act as agents using tools and external data.
Evaluation and testing
Honest measurement of what models can do: fair comparisons, uncontaminated benchmarks and testing against adaptive attacks.
Robustness and control
Defences against prompt injection, data poisoning and agent goal hijacking, and least-privilege permissions for the tools agents use.
Oversight and risk management
Technical evidence instead of declarations: how to demonstrate a system’s robustness and security under Article 15 of the EU AI Act and the Cyber Resilience Act.
Action plan
2026
- Bias Detection Under Article 4a of the EU AI Act: When Synthetic, Anonymised or Differentially Private Data Is Enough. Technical evidence options, a simulation study and a real-data replication
- Using Generative AI to Create Synthetic Medical Data for Patient Privacy Protection: A Review and an Empirical Evaluation of the Fidelity-Utility-Privacy Trade-offData package: Synthetic medical data and patient privacy benchmark (WHAS500, WDBC, FLCHAIN)
Planned for 2027
Topics of the technical reports and studies we plan to publish by the end of 2027. They will be published in English.
- Agent hijacking
- Fair comparisons
- Benchmark contamination
- Multi-agent risks
- Prompt injection
- Tool sandboxing
- Robustness evidence
- and many more
Articles from this area
Find articles on this topic in Research & Insights.