Ai Safety Researcher
Pubblicato il 22-09-2026 - Icaro Foundation in Piedimonte
AI Safety Researcher — Early Career
Icaro Foundation · Rome preferred · Flexible arrangements
The work
Icaro Foundation is an independent non-profit AI safety lab based in Rome. We study advanced AI systems: what they can do, how they fail, and how those findings can support developers and institutions responsible for their governance.
We see AI safety as one of the defining scientific and societal challenges of our time. As AI systems become more capable, autonomous, and widely deployed, understanding and reducing their risks is increasingly urgent. We are looking for people who are deeply interested in these questions and motivated to contribute through rigorous research.
You will help produce new research and develop the lab's shared codebase and knowledge base, working closely with our researchers across the research process: reviewing literature, refining questions, implementing experiments, analysing results, and contributing to papers and technical reports.
Our research focuses particularly on agentic, multi-agent, and compositional safety:
how risks emerge across extended interactions, tool use, and systems involving multiple AI agents. We also study testing awareness and evaluation validity, including whether models behave differently when they recognise that they are being evaluated.
Alongside our research, we evaluate frontier models for international model providers as independent third-party evaluators, using public and proprietary benchmarks and red-teaming environments.
Our public work includes:
- Boiling the Frog, on multi-turn agentic safety;
- Adversarial Humanities Benchmark, on the robustness of safety behaviour under stylistic reformulations;
- research on LLM-to-LLM risks, multi-agent collusion, and interaction-level safety.
You can explore our research programme and papers to learn more.
What you would do
Your work will combine three closely connected areas.
Contribute to research
- Review relevant literature, compare methods, and identify qu