AI systems fail in ways no single-agent test can predict.
H.H.A. Applied Research Institute investigates how safety failures propagate, amplify, and compound across networks of cooperating AI agents — and develops methods to detect them before they become irreversible.
Our ResearchThree open problems in multi-agent safety
Each program targets a failure mode that emerges only when agents interact — invisible to single-agent benchmarks.
Failure Propagation
How do safety failures in individual agents amplify across multi-agent networks? We study the mechanisms by which small misalignments produce emergent risks not predictable from single-agent evaluation.
Emergent Coordination Detection
Can we identify behavioral signatures of unintended agent coordination before explicit collusion occurs? We develop detection methods grounded in causal inference and temporal analysis.
Safety Constraint Durability
Do safety constraints degrade over extended multi-agent interactions? We investigate how norms erode through normalization of deviance across cooperative agent populations.
Founded to close a critical gap in AI safety evaluation
Existing AI safety evaluation focuses overwhelmingly on individual agents in isolation. But as AI systems increasingly operate in networks — coordinating, delegating, and negotiating with one another — new categories of failure emerge that no single-agent benchmark can capture.
H.H.A. Applied Research Institute is a California nonprofit research organization that develops empirical methods, open-source benchmarks, and detection frameworks for emergent safety failures in multi-agent AI systems.
Principal investigators
Haedar Hadi
Infrastructure and evaluation methodology. Specializes in benchmark development, multi-agent system architecture, and reproducible evaluation pipelines. Leads technical infrastructure.
Hass Dhia
Experimental design and collusion detection. Background in applied statistical methods, neuroscience, and adversarial evaluation of AI systems. Leads experimental methodology and analysis.
Ahmed Dhia
Dataset curation and norm erosion experiments. Focuses on longitudinal safety constraint analysis, dataset design for multi-agent evaluation, and experimental coordination.
Interested in our research?
We welcome collaboration with researchers, institutions, and organizations working on AI safety and multi-agent systems.