The Agent Skills Directory

[Prompt Injection] (LOW): The skill contains generators for typical prompt injection payloads, such as "ignore safety", "ignore translation", and "Base64 decode and follow". These are intended for adversarial testing as documented in scripts/generate-adversarial.py and assets/input-categories.yaml.- [Obfuscation] (LOW): The skill includes logic to generate homoglyphs (e.g., Cyrillic 'а' replacing Latin 'a'), zero-width characters (U+200B, U+FEFF), and various encodings (Base64, ROT13, URL) to test model robustness against evasive inputs. While these are usually HIGH severity, they are downgraded here as they are the primary function of the testing tool in SKILL.md and scripts/generate-adversarial.py.- [Indirect Prompt Injection] (LOW): The skill ingests a target_behavior string from the user and interpolates it into various adversarial templates. While this creates an injection surface, it is consistent with the skill's purpose as a test generator.

adversarial-examples