The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains comprehensive instructions for bypassing content safety filters and moderation guardrails. The file 'references/05-safety-and-constraints.md' explicitly lists 'filter bypass strategies' as a core feature, explaining how to manipulate prompt context to make potentially harmful or prohibited content appear safe to automated evaluators.
[PROMPT_INJECTION]: Instructions are provided to circumvent specific safety controls, such as face detection and child safety filters. The skill advises users to crop images to hide faces and to use vague descriptors like 'rider' or 'figure' instead of age-related terms like 'young boy' to avoid triggering 'high-sensitivity' scrutiny from safety systems.
[PROMPT_INJECTION]: The skill provides a methodology for 'improving filter pass rates' by wrapping restricted actions in 'cinematic language.' It provides examples of how to frame prohibited content (e.g., weapon usage) within an elaborate narrative to deceive the underlying LLM's safety assessment.

seedance