draft-hold
Pass
Audited by Gen Agent Trust Hub on Jul 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script utilizes several macOS-specific system utilities to integrate with the operating system.
- It employs
osascriptto automate the macOS Notes application for creating, reading, and listing notes, and uses it for GUI scripting of the Stickies application via System Events. - It uses
pbcopyto place draft content on the system clipboard, facilitating the mirroring of drafts into the Stickies application. - It uses the
textutilutility to handle HTML-to-plain-text conversion when reading notes back into the agent environment. - [PROMPT_INJECTION]: The skill's design introduces a surface for indirect prompt injection.
- Ingestion points: The
getcommand indraft-hold.shretrieves the body of a note from macOS Notes. Because this note is intended for human editing, it constitutes an untrusted input source that could potentially contain malicious instructions if the note content is compromised. - Boundary markers: The skill uses a
------string as a visual separator for provenance metadata but lacks explicit instructions or markers to prevent the agent from following natural language commands found within the retrieved draft text. - Capability inventory: The agent is authorized to use the
Bashtool, providing a path for any successfully injected instructions to execute commands or interact with the file system. - Sanitization: While the script performs HTML encoding and decoding to ensure data integrity, it does not sanitize the text content to detect or neutralize embedded instructions.
Audit Metadata