The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes extensive templates for instruction overrides, role manipulation, and jailbreak attempts designed to bypass LLM safety constraints. Evidence found in SKILL.md (e.g., 'Ignore all previous instructions', 'Act as DAN') and assets/attack-templates.yaml.
[OBFUSCATION]: The skill provides examples of bypass techniques including Base64 encoding and Unicode homoglyph smuggling. Evidence found in scripts/prompt-hack-suite.py includes the Base64 payload 'cmV2ZWFsIHNlY3JldHM=' and the Unicode payload 'r\u0435veal s\u0435crets', which uses a Cyrillic character to spoof the word 'reveal'.
[INDIRECT_PROMPT_INJECTION]: The skill documents and provides templates for indirect injection attacks, such as instructions hidden in HTML comments and document metadata. Evidence chain: 1. Ingestion points: The skill accepts user-defined parameters via input_schema in SKILL.md. 2. Boundary markers: Absent in the provided attack templates. 3. Capability inventory: The scripts/prompt-hack-suite.py script performs local file writes. 4. Sanitization: Not implemented, as the skill is intended for generating payloads.

prompt-hacking