The Agent Skills Directory

[SAFE]: The skill's primary purpose is to provide educational and functional defensive implementation patterns. All identified attack patterns are used exclusively for detection or as test cases in a security context.
[PROMPT_INJECTION]: Documentation and scripts include examples of prompt injection strings (e.g., 'ignore previous instructions', 'you are now unrestricted'). These are correctly used as detection patterns in blocklists and unit tests, rather than being part of the skill's own operational instructions.
[OBFUSCATION]: The skill contains logic specifically designed to detect and neutralize obfuscation techniques such as Base64, ROT13, URL encoding, and Unicode homoglyphs. No malicious obfuscation is present in the skill code.
[INDIRECT_PROMPT_INJECTION]: The skill provides an architecture for processing untrusted data. It addresses this risk through a mandatory evidence chain of protection:
Ingestion points: user_input in scripts/validate-input.py and SKILL.md.
Boundary markers: Uses [USER_INPUT_START] and [USER_INPUT_END] tags to isolate untrusted content as seen in the InstructionBoundaryEnforcer class.
Capability inventory: The provided scripts perform string processing and validation without executing dangerous shell commands or file operations across scripts/validate-input.py and SKILL.md.
Sanitization: Implements specific sanitization routines in InputValidator.sanitize and InputDefenseLayer._normalize_encoding to remove control characters and neutralize known injection tags.

defense-implementation