Long-form writing on AI agent security, decision-rights, reversibility-graded authority, and the architectural primitives underneath them.
Short canonical versions land first on third-party publications (newsletters, magazines, preprint servers). This repo is the long-form home: full version of each piece with every citation, every figure, every example that the short version had to drop.
- June 9, 2026 — Action-Class Authority for AI Agents: A Verification-Side Reference — Whitepaper v1.0 (~28 pages, 18 chapters). Four-class reversibility taxonomy, manifest-declared classification, worst-case chain rule. Anchored in OWASP AISVS C9.2.6 + C9.2.7 (proposed for 1.01). Standalone repo.
- June 3, 2026 — Per-Step Authorization Is Not Chain Authorization. The Worst-Case Action Class Primitive. — Five independent standards-track and executive-framework surfaces converging on a shared architectural primitive: manifest plus deterministic gate plus chain rule plus four-field audit schema.
- June 2, 2026 — What I Learned Contributing Across Five Standards Surfaces in Five Months — Maintainer dynamics, vocabulary, issue-before-PR, prior art, reference implementations across OWASP AISVS / SPVS / Cornucopia / GenAI ASI and CSA NHI.
- May 30, 2026 — Investigation Is Reversible. Actuation Is Not. The Architectural Floor for Agentic AI. — The read/write architectural fold as design primitive: where the gate lives, why the asymmetry is binary not gradient.
- May 22, 2026 — The Decision-Rights Plane: An Architectural Gap in AI Security — Why the bottom two layers of AI security lack the right primitive, and how to contribute to closing the gap across standards surfaces.
All articles in this repo are licensed CC-BY-4.0 unless explicitly stated otherwise. Quote, reuse, and translate with attribution.
Mayur Agnihotri. LinkedIn: linkedin.com/in/mayuragnihotri. GitHub: @Mayur021.
Reference implementations of the architectural primitives discussed in these pieces:
- aisvs-action-class-reference: manifest-declared action class and worst-case chain rule.
- nhi-runtime-decision-rights: non-human identity runtime decision-rights.