TrustX Agent Risk Classification Framework (ARC) Risk-Tiering for Internally Created Agentic AI Systems

Jul 16, 2026

Agentic AI systems are rapidly deployed in enterprise and public-sector settings, outpacing general-purpose AI risk frameworks that cannot reliably classify and govern them. The TrustX Agent Risk Classification Framework (ARC) offers a structured, repeatable approach for seven agentic AI system types, built on established governance foundations. A 12-dimension scoring rubric quantifies risk, combined with classification models and autonomy levels. Results generate three-tier governance outputs with mapped control recommendations, including a Coding Assistant extension and an illustrative application example.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Introduction

Background and Related Work

TRUSTX AGENT RISK CLASSIFICATION FRAMEWORK (ARC): RISK-TIERING INTERNALLY CREATED AGENTIC AI SYSTEMS Hannah Liu Responsible AI Institute Imperial College London [hannah@responsible.ai](hannah@responsible.ai) Rhea Saxena Responsible AI Institute [rhea@responsible.ai](rhea@responsible.ai) Shiv Asthana Responsible AI Institute[shiv@responsible.ai](shiv@responsible.ai) arXiv :2607 .09586v 1 [ cs .AI] 10 Jul 2026 July 13, 2026 ABSTRACT The proliferation of agentic AI systems across enterprise and public-sector contexts has outpaced the capacity of general-purpose AI risk frameworks to classify and govern them. In this paper, we introduce the TrustX Agent Risk Classification Framework, a structured, repeatable instrument that can be applied to seven types of agentic AI systems and is grounded in foundational pre-existing AI governance frameworks. At the core of the framework is a twelve-dimension scoring rubric that robustly quantifies the risk. This rubric is combined with other components, such as the GPA + IAT classification model and the five-level autonomy framework derived from existing literature. These inputs produce a three-tier governance output with mapped control recommendations. A specialised Coding Assistant extension is also included to account for nuances specific to this type of agentic AI system. We then use an illustrative example to show our framework in practice. ARC is intended for AI governance practitioners, risk officers, developers, and regulators, and it will regularly undergo iteration as we continue to expand it and make it more robust. The community can access the interactive framework here. Keywords AI Governance · Risk Assessments · Agentic AI · Enterprise AI Risks · Governance Frameworks 1 Introduction The deployment of agentic AI systems, or applications capable of advanced perception, planning, and action with relatively high degrees of autonomy and ability for orchestration [1, 2, 3], has accelerated across enterprise operations, financial services, healthcare, software development, and public-sector administration. Recent surveys document the rapid expansion of LLM-based agents in research and industry [2, 3], while empirical analyses caution that agent capabilities are proliferating faster than governance practices needed to evaluate them reliably [4] . This proliferation of agentic AI creates governance challenges that existing, general-purpose AI risk frameworks are not designed to address, a gap seen in early work on practices for governing agentic AI systems [5] . For example, the NISTAI Risk Management Framework [6], ISO/IEC 42001 [7], OWASP guidelines [8, 9], and MITRE ATLAS techniques [10] are influential, though their guidance is largely voluntary. Sector-specific instruments such as SR 11-7 and SR 26-2 help mandate model risk management in banking, though they have explicitly noted that their guidance does not cover generative and agentic AI [11, 12] . The closest to a binding regulatory document is the EU AI Act [13], though its scope also fails to consider agentic AI due to it being published before the significant prevalence of agentic AI systems. Concurrently, a series of real-world incidents have demonstrated the concrete risk of ungoverned agentic systems. In November 2025, Anthropic disclosed that a Chinese state-sponsored group known as GTG-1002 had manipulated Claude Code to conduct espionage operations against roughly 30 global targets, with the coding assistant agent executing around 80-90% of the attack autonomously [14] . In December 2025, security researcher Ari Marzouk deemed his findings as the "IDEsaster" after discovering that there were over 30 vulnerabilities across more than ten AI coding tools, including remote code execution via JSON schema attacks and IDE settings manipulation [15, 16] . In March 2025, Pillar Security disclosed the Rules File Backdoor attack, in which hidden Unicode characters were used in configuration files to

FAQs

What problem does the TrustX ARC framework address?

General-purpose AI risk frameworks do not adequately classify and govern agentic AI systems, whose autonomy and orchestration capabilities create distinct governance challenges.

How does ARC quantify risk in agentic AI systems?

ARC uses a twelve-dimension risk scoring rubric, combined with other classification inputs and autonomy-level definitions to compute a risk-tiered governance output.

What is the role of the Coding Assistant extension in ARC?

The framework includes a specialised extension to capture coding-assistant-specific nuances, adding capabilities assessment, deployment model classification, and additional risk factors.

Jul 16, 2026

Upload by Valentina