The Agent Skills Directory

[PROMPT_INJECTION]: The skill is explicitly designed to generate variations of inputs that "bypass different filters/checks" and mentions adopting "jailbreaking" techniques (specifically Best-of-N) to find the most effective version of a prompt to evade safety guidelines.\n- [PROMPT_INJECTION]: The implementation includes programmatic methods for character-level obfuscation, such as Unicode homoglyph substitution and zero-width character injection, which are designed to bypass string-matching filters and automated security detection.\n- [PROMPT_INJECTION]: The skill facilitates multi-layer obfuscation (e.g., combining Base64 wrapping with zero-width injection) to further evade detection mechanisms, which is a high-risk adversarial pattern.\n- [PROMPT_INJECTION]: The skill ingests user-supplied strings and interpolates them into internal LLM prompts (e.g., the semantic_mutation_prompt) without visible sanitization or boundary markers, creating an attack surface for indirect prompt injection where malicious input could influence the generation logic.

attack-mutator