unix-macos-engineer

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEPERSISTENCECOMMAND_EXECUTIONCREDENTIALS_UNSAFEPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PERSISTENCE]: The skill provides detailed templates and instructions for creating macOS launchd agents and daemons in references/launchd-patterns.md. This allows for scripts to be executed persistently across reboots, user logins, or based on file system triggers.\n- [COMMAND_EXECUTION]: Extensive documentation is provided for powerful macOS-specific and Unix command-line utilities. This includes tools like osascript for executing AppleScript, defaults for system preferences manipulation, and launchctl for service management.\n- [CREDENTIALS_UNSAFE]: The reference guides in references/macos-commands.md document the use of the macOS security utility to read from and write to the system Keychain (e.g., security find-generic-password). While a standard administrative task, this provides a surface for the agent to interact with sensitive stored credentials.\n- [PRIVILEGE_ESCALATION]: The skill documents the use of sudo for several system-level operations, such as rebuilding Spotlight indexes and flushing DNS caches. It also provides shell scripting patterns to detect if a script is running with root privileges via the $EUID variable.\n- [EXTERNAL_DOWNLOADS]: The skill includes instructions for using Homebrew (brew install) to download and install third-party packages and applications.\n- [INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect injection as it processes various system inputs that could be influenced by untrusted data. \n
  • Ingestion points: Use of pbpaste (clipboard), mdfind (file metadata), reading arbitrary file contents, and defaults read. \n
  • Boundary markers: The skill explicitly instructs the agent to use defensive scripting patterns like set -euo pipefail and proper variable quoting. \n
  • Capability inventory: Includes full shell access, file system read/write, network operations via curl, and persistence via launchd. \n
  • Sanitization: No specific sanitization logic is implemented in the skill body; instead, it relies on the agent following the provided defensive coding guidelines.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 01:25 PM