The Agent Skills Directory

[COMMAND_EXECUTION]: The file references/command-blacklist.md contains a comprehensive list of high-risk shell commands including rm -rf, nc -e /bin/sh, and chmod 777. These entries are strictly data definitions for a security filtering engine designed to intercept and prevent malicious command execution by users, rather than executable code within the skill itself.\n- [PROMPT_INJECTION]: The skill defines a set of detection signals for prompt injection in references/static-rules.md and SKILL.md, covering both English and Turkish patterns (e.g., 'ignore previous instructions', 'DAN mode'). These patterns are documented as detection signals for the filtering layers and are not used to subvert the agent's instructions.\n- [INDIRECT_PROMPT_INJECTION]: The architecture includes an LLM-based security auditor ('LLM Judge') that processes message content. This component represents a potential vulnerability surface if an attacker attempts to influence the judge's prompt instructions.\n
Ingestion points: User messages and context are ingested into the LLM Judge prompt in SKILL.md.\n
Boundary markers: The prompt template uses simple bracket interpolation without multi-modal delimiters or instructions to ignore embedded commands.\n
Capability inventory: The skill can block messages, update security caches, and synthesize new rules.\n
Sanitization: There is no explicit description of input sanitization for the message data before LLM processing.

adaptive-guard