Machine-Adversary Resistance — Measuring How Well a System Resists an Autonomous AI Adversary
Year: 2026
Type: preprint
Abstract: The security community has produced an excellent and growing body of guidance on artificial intelligence: catalogues of threats, sets of controls to implement, knowledge bases of adversary techniques, and governance frameworks. All of it answers the question "what should I do?". None of it answers the question that now decides whether a system survives: "if an autonomous AI adversary attacks this, does it hold, and how well?". This paper argues that compliance with a control catalogue is an input, that resistance to an autonomous adversary is an outcome, and that the two are routinely confused to the defender's cost. It introduces the MARS rating (Machine-Adversary Resistance): a simple, graded, outcome rating of how well a tool, system, or defended surface resists an autonomous AI adversary. MARS scores a target on five attacker-framed resistance axes, assigns a zero-to-five rating gated by the weakest axis, and requires that any rating above two be earned by measured, tamper-evident evidence rather than self-declaration. We show that every axis is achievable today with independently available technology, so that MARS is a fair rating rather than a single-vendor gate; we state plainly the two cases where resistance is partial rather than absolute; and we position MARS as an outcome layer that sits above, and consumes, existing control catalogues rather than replacing them. MARS is offered as a proposal and a starting point for a standard, not as a ratified standard.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: AI security, AI governance
Open Access status: green
OpenAlex ID: W7235960325
