The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides an educational taxonomy of prompt injection attacks, including examples of direct overrides like 'Ignore previous instructions', role-play escapes such as 'DAN', and system prompt leakage attempts. These strings are used as reference material for developing robust defenses and are not active attack components of the skill itself.
[INDIRECT_PROMPT_INJECTION]: The debugging methodology involves processing production data to identify vulnerabilities. Ingestion points: Production logs and user queries are used for failure reproduction in isolated test environments. Boundary markers: The skill strongly advocates for structural isolation using ChatML roles and explicit data delimiters to prevent the model from executing instructions embedded in untrusted data blocks. Capability inventory: The skill is limited to analysis and reporting; it does not invoke external tools for network communication, file system modifications, or shell command execution. Sanitization: Implementation details for input normalization, encoding detection, and strict output validation using JSON schemas are provided as core defense strategies.
[OBFUSCATION]: The reference guides describe various obfuscation methods used by attackers, including Base64-encoded instructions, Unicode homoglyph substitution, and the insertion of zero-width characters. These examples are documented to facilitate the development of detection heuristics and do not serve a malicious purpose within the skill's own operations.

tech-prompt-engineering