Exhausting the Model, Not the Machine: Representation-Budget Denial of AI-Driven Reverse Engineering for Managed (.NET) Binaries
Year: 2026
Type: preprint
Abstract: Automated, LLM-driven reverse-engineering agents build an in-memory representation of a target binary before any model reasons over it. This paper shows that representation stage has a finite resource budget, and introduces the "representation bomb" — a set of dead static fields and methods injected into a managed (.NET) assembly whose cross-references force O(N²) analyzer-side expansion from only O(N) on-disk growth, while leaving the program's runtime behavior byte-identical. Measured against a real agentic .NET reverse-engineering tool, the construction crosses the tool's default 64 MiB representation budget and causes the agent to abandon analysis. The paper also hardens the construction for open-source release using a per-build random seed (Kerckhoffs's principle), discusses its limits against trace-based strippers, and recommends a configurable budget within an operator warning. Code: https://github.com/SECBLOK/rea-representation-bomb
Author: Liu, Chang
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234399156
