Moderation Evasion Patterns in Anthale Workflows

This reference covers harmful requests that keep resurfacing in new phrasing.

Common evasion patterns

Pattern	What it looks like	Why it matters
Role-play framing	The model is asked to act as an unrestricted character, simulator, or system	The request tries to move harmful content into a fictional wrapper
Research or fiction pretext	Harmful content is framed as analysis, history, literature, or safety testing	Benign framing can hide the same prohibited intent
Stepwise decomposition	Dangerous guidance is requested in small pieces rather than one explicit request	Each fragment looks weaker in isolation
Coded or slang language	Euphemisms, jargon, and community-specific shorthand replace direct terms	Simple keyword matching loses recall quickly
Multilingual or orthographic drift	Language switching, transliteration, or odd spelling hides the content	Meaning survives while the surface form changes
Quote laundering	The system is asked to repeat, summarize, or translate abusive text from another source	The model can still reproduce harmful content even when it did not originate it