Deep Dives
Long-form investigative reports on AI safety. Expanded research access is planned after entitlement services are deployed.
Twelve AI Failures That Shaped Modern Risk Thinking
A concise field guide to recurring failure modes.
From legal hallucinations to autonomous vehicle failures, the same risk patterns appear across domains.
The Frontier Model Evaluation Landscape in 2026
A deep analysis of how frontier models are evaluated — and why the evaluation gap is the real bottleneck.
From HELM to SWE-bench, the evaluation ecosystem is fragmented. This report maps the benchmarks, identifies the gaps, and proposes what a unified safety evaluation should look like.
AI Agent Safety: The New Frontier of Risk
When models gain tools, the risk surface explodes. This report examines the safety challenges of agentic AI systems.
Agentic AI is the most consequential deployment trend of 2026. It is also the least understood from a safety perspective. This report examines what makes agents different — and what safety frameworks are emerging to address them.