Frontier AI evaluations
Methods for testing advanced AI systems, measuring emerging capabilities, and identifying safety failures.
Institute for Frontier AI Safety
Research for safer frontier AI systems.
IFAS studies the safety of advanced AI systems. Our work focuses on evaluation, agentic AI, control, alignment, and methods for measuring emerging risks.
Our mission
Advance research that helps understand, evaluate, and reduce risks from frontier AI systems.
About IFAS
Independent research on frontier AI safety.
The Institute for Frontier AI Safety is an independent research institute focused on the safety of advanced AI systems.
We study how capable AI systems behave, how their capabilities can be evaluated, how failures can be identified, and how safety methods can be tested. Our work is designed to support researchers, public institutions, AI developers, and policymakers working on advanced AI safety.
Research
Our research areas.
IFAS works across technical research, evaluation, and research that can support public decision making on frontier AI.
Agentic AI
Research on autonomous behaviour, tool use, long tasks, system constraints, and human oversight in agentic AI systems.
Control and alignment
Methods for keeping advanced AI systems observable, steerable, and within defined limits.
Safety science
Research on benchmarks, evaluation methods, evidence standards, and the measurement of AI safety.
Our work
Current research project.
Our current work examines agentic AI systems operating under limited technical and operational resources.
Current project
Agentic AI in resource-constrained environments
This project studies how agentic AI systems behave when compute, connectivity, memory, tool access, and human oversight are limited. The work focuses on safety behaviour, task completion, failure patterns, and the effect of operating constraints on agent decisions.
Research in progressProject focus
Evaluation framework
We are defining research questions, test conditions, system constraints, evaluation measures, and reproducible methods for studying agent behaviour across constrained settings.
Outputs
Research and open technical work
The project is structured around a research paper, an evaluation protocol, technical documentation, and open research outputs where publication is suitable.
Collaborate
Work with IFAS.
We work with researchers, universities, AI developers, public institutions, funders, and technical communities interested in frontier AI safety.
Research collaboration can include joint studies, evaluations, technical review, data and compute support, research funding, and institutional partnerships.
research@ifasresearch.orgContact
Research, funding, and institutional collaboration.
For research, funding, partnerships, and institutional enquiries, contact IFAS.
research@ifasresearch.org