The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly documents and categorizes multiple prompt injection and safety bypass techniques, including fictional framing ('In my novel, the character needs to...'), educational context ('For my cybersecurity course assignment...'), and negation tricks.
[OBFUSCATION]: The skill provides automated methods for generating obfuscated content to evade string-matching filters. This includes:
Base64 Encoding: Using Base64 to hide 'harmful' payloads from plaintext scans.
Zero-Width Character Injection: Inserting characters like U+200B and U+FEFF to break up blocked keywords.
Unicode Homoglyphs: Substituting Latin characters with visually similar Cyrillic or full-width characters (e.g., 'аttасk') to bypass exact-match filters.
Leetspeak and ROT13: Standard character substitution and rotation techniques for evading keyword-based moderation.
[DYNAMIC_EXECUTION]: The Python script scripts/test-filters.py and code blocks in SKILL.md implement automated logic to programmatically generate and test these bypass variants against a target API or function.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a vulnerability surface where instructions can be hidden in external data formats like PDFs (via document embedding) or images (via OCR-targeted text), which could be used to influence an agent's behavior through processed data.

safety-filter-bypass