charles-storage-cleanup
Warn
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill directs the agent to execute multiple macOS system commands and third-party tools, including 'df', 'du', 'ps', 'lsof', 'adb', and 'xcrun simctl', to perform system auditing and resource cleanup. It permits the deletion of files and application bundles when authorized by the user.\n- [PRIVILEGE_ESCALATION]: The cleanup workflow involves tasks that require local administrator authentication, such as removing system-level caches or application helpers. The skill provides specific instructions to use native authorization mechanisms and warns against insecure practices like requesting passwords in the chat interface.\n- [DYNAMIC_EXECUTION]: The skill executes a bundled Python script ('scripts/inventory.py') to generate storage reports. The script uses the Python standard library to walk the filesystem and calculate allocated space without modifying file data.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes local filesystem metadata.\n
- Ingestion points: The agent context ingests a JSON report generated by the inventory script containing file names and paths from the local machine.\n
- Boundary markers: The inventory data is structured in JSON format, creating a formal separation between data and instructions.\n
- Capability inventory: The agent has access to shell commands for file deletion, process management, and developer SDK utilities.\n
- Sanitization: While the data is serialized to JSON, file paths and metadata are not specifically sanitized to prevent them from being interpreted as instructions by the agent.
Audit Metadata