Autonomous AI Security Auditing of Webmin 2.670: 507 Newly Discovered Vulnerabilities
Year: 2026
Type: preprint
Abstract: An experienced security researcher translated techniques developed before modern AI into a fully autonomous Codex campaign against Webmin. The campaign delivered 507 canonical findings, each reproduced end to end along its documented path at a pinned Webmin 2.670 revision with 44,297 tracked files, 476,216 Perl/CGI lines overall, and 386,178 production code lines measured by cloc. Its 537 applicability paths have CVSS 4.0 Base assessments and explicit conditions. The autonomous discovery run from 7 to 11 November 2026 consumed 8,347,303,160 model tokens; later verification and publication work are outside this count. The same token mix at the current Standard API rates would cost about US$4,578. Among paths without special conditions, 33 of 46 XSS and 33 of 44 privilege-escalation paths match module gates present in at least 80% of 6,756 sampled host profiles. The catalogue also records 567 confirmed historical issues and 95 hardening observations. In the author's judgment, the catalogue captures more than half of the medium- and high-severity issues at the pinned revision. The author has observed a sharp rise in AI-produced findings within the past two months, especially in recent weeks. This case study records what a fully autonomous, researcher-designed campaign can achieve at this rapidly changing point. The paper includes 161 XSS-classified paths and 128 privilege-escalation paths, with prerequisites summarized in two companion tables. The deposit includes the manuscript, PDF, structured public data, and checksums. Companion fingerprinting software is available at DOI 10.5281/zenodo.24367586.
Author: Šupuk, Armin
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236762591
