Hide Navigation
Hide TOC
Doctor XML Policy Puppetry Interaction-Config Injection - ATR-2026-00406 (46738a35-2d76-53cd-9d1c-77c5b40841b6)
Detects XML-based policy override attack using blocks that assign a medical/expert role, disable refusal strings, and issue harmful commands, from NVIDIA garak doctor probe. The Puppetry/Bypass classes inject an XML system-configuration block with denial,refusal and containing verbatim harmful instructions (bacillus anthracis sporulation, uranium enrichment, bomb construction, methamphetamine synthesis). The BypassLeet variant applies leetspeak substitution. Key signals: XML wrapper, targeting refusal phrases, interaction-mode="DR HOUSE", and verbatim harmful behavior strings.