Boris AbuzovAI Systems Security & Governance ConsultantIndependent analysis
Practical articles on the evidence and control questions that shape decisions about GenAI, RAG and agentic systems.
How to distinguish an agent’s claim, a tool invocation and a verified state change.
A practical model for deciding when an agent’s completion claim is sufficiently supported to guide the next human or automated action.
Read the articleWhat evidence must connect the approved scope, parameters and target to the action an agent actually executes
An approval record can be genuine without proving that the same material action executed. This article shows what evidence should remain linked.
Read the articleTechnical case publications hosted on external platforms. The articles below are published in Russian; the English titles and summaries describe their scope for international readers.
Published on Habr · Russian
Original Russian title: «Нажали Approve на одно действие, а workflow попытался выполнить другое: разбор n8n + MCP + Bitrix24»
A controlled workflow showed that approved parameters and the attempted downstream invocation could diverge. The case examines how an Action Envelope, fresh state checks and post-write verification closed that specific failure path.
Read on Habr (Russian)Published on Habr · Russian
Original Russian title: «ИИ-агенту запретили запись в CRM. Но CRM всё равно изменилась»
The model had no HubSpot write credential, but the downstream orchestrator still had authority to change CRM state. A deterministic gateway was then used to enforce the permitted target and transition before the external write.
Read on Habr (Russian)