> ## Documentation Index
> Fetch the complete documentation index at: https://anthale.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Moderation Evasion Patterns in Anthale Workflows

> Classify the common ways abusive requests are disguised to get past moderation rules across prompts, transformations, and indirect model 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 |

## Next steps

From here, go to [Prompt Injection Protection](/docs/learn/guardrails/prompt-injection) or return to [Content Moderation](/docs/learn/guardrails/content-moderation).