Research posts from Redwood Research on catastrophic AI risk, focused on AI control, monitoring of agentic models, and red-teaming of safety systems. Readers get technical write-ups of papers, experiments and engagements with AI developers, including methods, results and open questions.

A security leader's newsletter on AI agent security, introducing a framework called Wingbeat Radius for measuring how far untrusted input can travel through AI agents. It analyzes real incidents such as the Cline supply chain compromise and an AI sandbox escape, and covers published vendor controls for limiting agent reach.

Analysis of AI agent security, cost and compliance, drawing on benchmarks, industry data and the author's experience building and securing agent platforms. Each issue argues a specific position, such as treating tool surfaces as dependency trees or the context window as a budget.

Field notes and essays on securing and governing AI systems, especially agents, adapted from the author's book and conference talks. Readers get conceptual frameworks and practical questions for deciding what AI should be trusted to do, and how to test and monitor it.