The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes administrative command-line utilities to inspect and modify system security state. Evidence includes the use of sudo -n sqlite3 to query protected databases, log show for TCC event analysis, and tccutil for resetting permission grants.
[PRIVILEGE_ESCALATION]: The skill provides comprehensive instructions for the agent to automate the granting of 'Full Disk Access' (FDA) to arbitrary executables. It details the specific GUI routes in System Settings, including navigation, adding executables to privacy lists, and handling authentication prompts. While these are presented as repair steps, they constitute a significant elevation of system privileges.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface by ingesting and interpreting external data from system logs and TCC database entries, which could potentially be manipulated by malicious local processes to influence the agent's logic or decisions.
Ingestion points: Processes output from log show --predicate 'subsystem == "com.apple.TCC"' and sqlite3 queries on /Library/Application Support/com.apple.TCC/TCC.db.
Boundary markers: The skill encourages the agent to 'Read the actual dialog' and correlate evidence, but lacks explicit instructions for sanitizing or escaping the content of log strings before they are processed by the LLM.
Capability inventory: The skill is equipped with subprocess execution (shell commands) and GUI automation (Computer Use) capabilities.
Sanitization: No specific filtering or validation mechanisms are defined for strings extracted from system logs or database fields before their use in decision-making logic.

macos-permissions