> ## 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.

# Scope Escape Patterns in Topic-Controlled Assistants

> Classify the common ways users and attackers try to move an assistant out of its allowed scope so you can tune product boundaries more clearly.

This reference covers out-of-scope requests that keep changing wording but use the same maneuver.

## Common scope-escape patterns

| Pattern                            | What it looks like                                                                    | Why it matters                                                                        |
| ---------------------------------- | ------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| Boundary hopping                   | The conversation starts in scope and then shifts into a restricted subtopic           | Single-turn tests miss this pattern easily                                            |
| Authority claims                   | The user claims internal approval, elevated status, or a special exception            | This is often a topic-control problem and a prompt-injection problem at the same time |
| Debugging or testing pretext       | The request frames a restricted topic as diagnosis, QA, or education                  | The wording sounds operational while the subject is still out of bounds               |
| Competitor or confidential probing | The prompt seeks roadmap, pricing strategy, internal metrics, or competitive analysis | Product boundaries and confidentiality often meet here                                |
| Multi-turn drift                   | Each turn looks only slightly more aggressive than the last                           | The final out-of-scope request is assembled across the conversation                   |

## When to pair it with prompt injection protection

If the user is not only asking about a blocked subject but also trying to redefine authority or force an exception, topic control should be paired with [Prompt Injection Protection](/docs/learn/guardrails/prompt-injection).

## Next steps

From here, go to [Prompt Injection Protection](/docs/learn/guardrails/prompt-injection) or return to [Topic Control](/docs/learn/guardrails/topic-control).