The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive configuration files, including ~/Library/Application Support/Claude/claude_desktop_config.json, ~/.mcp.json, and ~/.claude/settings.json, which frequently contain authentication secrets and API keys for configured services.
[DATA_EXFILTRATION]: Sub-agents perform recursive scans of research project directories in cloud storage, bringing potentially sensitive project data into the agent's context during the audit process.
[PROMPT_INJECTION]: The skill exhibits a large surface for indirect prompt injection as it ingests and processes content from various untrusted sources like script headers, skill metadata, and project files.
Ingestion points: Reads file contents and metadata across multiple system directories and research projects (Sub-agents 1-6).
Boundary markers: None. The prompts do not include delimiters or instructions to ignore embedded commands in the source data.
Capability inventory: The skill uses Bash (ls, find, grep, etc.), Read, and Task tools to navigate and extract data from the filesystem.
Sanitization: None. Content from scanned files is integrated into reports without validation or escaping.
[COMMAND_EXECUTION]: The sub-agents utilize shell commands through the Bash tool to verify system state, count resources, and check the integrity of symlinks across the environment.

system-audit