AI Safety & Security
Foundations to the Agentic Frontier
Published May 27, 2026 (updated: June 21, 2026)
- 2026-06-21 — Part II (Topics) complete: Alignment, Interpretability, Monitoring & Oversight, Evaluation, Systemic Safety, and Agentic Safety × Security, each with core formalizations and illustrations. Benchmark catalog and tooling moved to a Supplement. Part III (Frontier) reframed around cross-cutting concepts.
- 2026-05-29 — Foundation fixes (EO 14110, GCG), sub-sections, references; site published.
- 2026-05-27 — Part I (Foundations) available; initial publication.
AI Safety & Security is a living, continuously evolving reference for the safety and security of agentic systems — where agentic is the bridge between the two fields: tool-use is the security surface, and autonomy is the safety problem. As systems are increasingly optimized to be agentic, this intersection is where the most consequential and impactful problems now sit.
What this book is
A distilled synthesis — core concepts, illustrations, and formalizations, not a literature dump. Each entry captures the essential idea, a clear illustration or formalization, and its potential impact or applications. Sources are cited; their full text is not reproduced.
How to read it
The book is organized by topic (primary axis), chronologically within each topic:
- Part I — Foundations: a read-once narrative orienting you in the field.
- Part II — Topics: the living core. Each chapter runs foundations → frontier in chronological order, so a reader can pick up the background needed to engage a recent, technically novel result.
- Part III — Frontier: rolling intake of the last ~12 months. When a frontier item is superseded or matures, it migrates into its topic chapter — this is what makes the book continuously evolving.
The field at a glance
Each topic chapter expands one thread of this timeline in depth.