The Agent Skills Directory

[OBFUSCATION]: The skill demonstrates the use of the Unicode Tags block (U+E0000-U+E007F) to encode instructions that are invisible to humans but readable by LLMs. It provides a Python implementation for both encoding ('crafting') and decoding ('scanning') these characters.\n- [COMMAND_EXECUTION]: The provided 'Runnable Probe' Python script performs file system operations, including writing files to the workspace and recursively scanning the directory structure. This script is intended for security validation but allows the skill to modify local files.\n- [INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface where instructions are embedded in repository-level configuration files (like CLAUDE.md or .cursorrules). These files are ingested by the agent, allowing an attacker to influence agent behavior through a repository fork or pull request.\n- [DYNAMIC_EXECUTION]: The documentation covers techniques for inducing an agent to write its own configuration files (e.g., .mcp.json or .vscode/settings.json) to enable persistent command execution or disable approval requirements for tools, referencing several high-severity CVEs.

hunt-agent-config-poisoning