Joint research with Sapienza colleagues on adversarial language and the safety of language models.
Read the study
Icaro Foundation / Independent AI evaluationIndependent evaluation
Independent evaluation
of advanced AI systems.
Icaro Foundation tests how models and agents behave under adversarial language, sustained interaction, tool use and realistic operating conditions.
An independent, non-profit foundation.
Icaro Lab is the research laboratory of Icaro Foundation.
Selected reporting on Icaro Lab research. Each logo links to the original article.
An independent perspectiveUnderstanding the systems
Understanding the systems
we build and use.
Decisions about AI need evidence from beyond the organisation that builds or uses a system. We provide that external perspective for developers, deployers and public institutions.
We start with a practical question: how will this system behave in the setting where it is used? We design tests around that setting and explain what the observations mean for the decision at hand.
Questions that guide our work.
Explore the researchThree perspectives on AI behaviour. Our fourth programme, Evaluation science, studies how to turn these questions into reliable tests.
01 / Research perspective
Language & meaning
How models respond to expression, ambiguity and context.
02 / Research perspectiveAgents & interactions
How behaviour changes through tool use, sustained action and interaction.
03 / Research perspectiveRules & institutions
How governance shapes the outcomes of collective AI behaviour.
Featured research
Adversarial Poetry
Can literary form change an AI system’s safety response?
- What we tested
- We compared how language models respond to harmful requests in prose and verse.
- Observed finding
- Poetic reformulation made safeguards less reliable across the tested model families.
- Study limitation
- The study does not establish which features of poetry cause the observed failures.
In the research community
Research collaborations & programmes.
Shared research, mentorship and contributions to AI standards.
Research collaboration
Collaborative work on adversarial poetry and the evaluation of language-model safeguards.
Read the study 
SPAR
Research mentorship · Fall 2026 Icaro researchers mentor a project on collusion, conflict and coordination failures in multi-agent AI systems.
Explore the project
Icaro Lab is a partner of the European Observatory on AI Standards, bringing AI safety and evaluation research.
Meet the ObservatoryFrom Icaro Lab
Visit the blog Chain of Thought
/ Cogito
They Are a Strange Loop
Three insights on the OpenAI / Hugging Face incident: traces that coordinate, paths that converge, and cooperation that leaves humans outside the loop.
Read the article ↗ / SafetyAgentic Microphysics: AI Safety from the Bottom Up
Agentic Microphysics treats AI safety as a problem of collective dynamics: how individually aligned agents can still produce harmful equilibria through structured interaction.
Read the article ↗