VFA: Target-Bound Evidence for Multi-Agent Vulnerability Investigation
Year: 2026
Type: preprint
Abstract: An agentic vulnerability report can preserve plausible reasoning while losing the identity of the bytes on which it depends. A decompiler address may be cited as though it were an original file offset, or several descriptions of one input-to-sink path may be counted as separate vulnerabilities. In a retained VFA scan of a PE executable, four Finding entries described input reaching the same copy operation; a subsequent audit grouped them as one path while retaining both input variants. This paper presents VFA, a vulnerability-investigation system that binds inspection records and role submissions to a verified local target and requires citations to match retained locations. Investigators submit candidates; a reviewer in a separate work context and a proof planner examine each candidate before adjudication; a coordinator records the resulting status and gaps. In seven selected primary attempts, 40 Finding entries were grouped into 34 paths, 26 of which an AI-assisted, author-side audit judged statically supported under stated conditions. A separate complete ASAR companion contributes two more entries but no new path. Across the eight selected exports, 314 historical citation locations and retained Evidence digests matched. In a separate preregistered current-code checker study, 22 valid inputs were accepted, 63 malformed inputs were rejected, and one semantic limit was recorded as a gap across seven route fixtures; all 86 outcomes matched the frozen oracle. None of the 42 PoCs was run. These observations establish traceability and tested checker behavior on selected records, not detection accuracy, runtime exploitability, or superiority over other agents.
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225448373