Architecture
How is the system structured, and how do information, authority and responsibility move through it?
About Boris Abuzov
I am an independent AI Systems Security & Governance Consultant specializing in Agent Authorization, Runtime Controls & Execution Evidence. I conduct Independent AI Risk Reviews of consequential AI systems, supporting decisions about production reliance, client handover, scaling, increased autonomy, and privileged or write actions.
Why This Practice Exists
My current work brings together experience from several distinct stages of my career.
I began by working with technical products and international manufacturers before moving into commercial leadership roles. This work required technical, operational, and business considerations to be evaluated together rather than in isolation.
I then moved into professional work in information security and cybersecurity. Alongside that work, I strengthened my technical foundation through networking studies aligned with the CCNA curriculum and additional training in information security and cybersecurity fundamentals. This shaped how I think about system architecture, access, trust boundaries, permissions, and operational risk.
I subsequently founded and led a specialized B2B company supplying industrial equipment. Running the company meant bringing together technical products, customer requirements, contractual obligations, financial consequences, and direct responsibility for delivery and outcomes.
As AI systems became relevant to real operational and business decisions, these different strands of experience converged. Agentic and tool-enabled AI systems brought a particular problem into focus: the gap that can exist between what a system is intended to do, what it is authorized to do, and what it can actually do—or cause—at runtime.
This raises practical questions. Does an action remain within the authority granted to the system? Do runtime controls still hold at the point of execution? Does the executed action correspond to the action that was authorized? And does the available evidence establish what the system actually did and what changed as a result?
These questions led to my current practice. I review consequential AI systems by connecting architecture, risk, operational consequences, and controllability, with a particular focus on agent authorization, runtime controls, and execution evidence.
My role is to provide an independent, evidence-based assessment of whether the system’s architecture, controls, and execution evidence support the decision being considered. This work complements the responsibilities of security, legal, compliance, and engineering teams; it does not replace them or constitute certification.
Current Practice
My practical work has included a bounded review of an early-stage agentic AI system. Documentary analysis was extended with stakeholder explanations, demo-based observations, and a remediation and validation plan.
My current technical work examines where operational authority resides across models, orchestrators, tools, APIs and downstream systems; whether intended controls hold at runtime; and what evidence reconstructs the action and resulting state.
The practice is supported by the AI Systems Risk & Evidence Lab, which develops synthetic cases, controlled experiments and reusable evidence workflows.
The Lab includes a RAG Risk & Evidence Harness for controlled examination of retrieval behavior, source boundaries, context construction and traceability.
How I Approach AI Risk
I review AI as part of a wider operating system of models, data, retrieved information, orchestrators, permissions, tools and APIs, human decisions, downstream systems and business processes. I examine where authority to act actually resides, how controls bind to consequential actions at runtime, and what evidence supports reconstruction of system behavior.
The overall review perspective follows five connected questions:
Risk
Which technical or operational mechanisms could produce incorrect, unreliable or insufficiently controlled behavior?
Consequences
How could that behavior affect users, operations, data, contractual commitments or business decisions?
Controllability
Can incorrect actions be prevented, unexpected behavior detected, the system interrupted when necessary and operations restored after failure?
Decision Conditions
What should be corrected, tested, observed or documented before the proposed next step has sufficient support?
At the level of an individual finding, these questions are developed through a more detailed structure connecting the mechanism, supporting material, consequences, existing strengths, control gaps, remediation and acceptance evidence.
Evidence Before Confidence
I distinguish the source and conditions of information from how it was reviewed and what it can support.
A documented control does not prove implementation. A demonstration may show behavior in one scenario without establishing representative operational performance. A controlled experiment may demonstrate a mechanism without proving production readiness.
The level of confidence should not exceed the information available.
Technical tools can produce traces, observations and test results. They do not make the final judgment. I conduct the review directly and remain responsible for connecting the evidence to consequences, controllability and decision conditions.
Who I Work With
I work with AI product companies, software teams, integrators and professional-service firms that need an independent view before deployment, client handoff, scaling or increased autonomy.
A review may address the wider system or a specific workflow, component, control question or proposed change.
Independence and Professional Boundaries
I do not sell an AI platform, implementation package or automatic risk score.
The review is structured around the decision being considered, the available evidence and the conditions requiring clarification. Its purpose is not to justify a predetermined outcome.
An Independent AI Risk Review is not a formal audit, certification, legal opinion, compliance approval, penetration test or red-team exercise.
Conclusions remain limited to the reviewed system, workflow, environment, information and decision context. Responsibility for the final decision remains with the relevant system owner.
More detailed limitations are available on the Professional Boundaries page.
Discuss Your AI System
Begin with a short overview of:
- what the system or workflow does;
- its current stage;
- the decision, change or next step being considered;
- the main uncertainty you would like to clarify.
You do not need to choose the final review route before making contact.
Detailed technical materials are not required for the first message.
Please do not submit passwords, credentials, API keys, sensitive personal data, confidential documents, unredacted logs, production traces, source files or archives in your initial email.