The Expanding Perimeter: A Maturity Model for Agentic AI Attack Surfaces
Enterprise security programs are built on the assumption that defended systems stay fixed between periodic reviews. Autonomous AI agents break that assumption because they decide at runtime which tools, peers, and data to access, and so redraw their own attack surface. Existing maturity instruments, including the Capability Maturity Model, COBIT-based assessment, and the NIST Cybersecurity Framework, were built to grade human-run practices that change slowly. Newer agent-specific governance models treat adaptive control as the top of a single scale rather than as an independently scored dimension. Two documented incidents, an external agent that reportedly reached McKinsey’s production database within about two hours, and a zero-click Microsoft 365 Copilot flaw that enabled data exfiltration through a single email, show that periodic assessment cannot keep pace with a system that redraws its own scope in the moment. This paper proposes a four-dimension, five-level maturity model comprising Visibility, Containment, Accountability, and Adaptability, each anchored to an observable test rather than a self-reported survey response. Applied to a composite financial advisory firm, the model scores each dimension independently rather than averaging them, surfacing an imbalance (strong detection, weak prevention) that a composite score would conceal, and yielding a prioritized remediation order that supports business sustainability. The framework is presented as a structured self-assessment instrument, with a defined pathway for expert review, inter-rater testing, and pilot evaluation. The model has not yet been validated against real organizations, and its Adaptability dimension remains constrained by the current state of governance tooling.
Authors
- Ganiyu Oladimeji (ORCID: https://orcid.org/0000-0002-3684-7428)
Institutions
- Moshood Abiola Polytechnic (NG)
Publication Details
- Published
- 2026-11-20
- DOI
- https://doi.org/10.68607/jaisb.v1i1.25
- Primary Topic
- Information and Cyber Security
- Type
- article
- Field-Weighted Citation Impact
- 0.00