officialese-penman
Warn
Audited by Gen Agent Trust Hub on Jul 31, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The
SKILL.mdfile contains instructions for the agent to invoke the shell to run the bundledscripts/check_placeholders.pyscript for document validation. - [DATA_EXFILTRATION]: The script
scratch/parse_html.pycontains hardcoded absolute paths pointing to a specific local user directory (/Users/zhangxin/). These paths reference internal data structures of the Antigravity IDE (.gemini/antigravity/brain/), which could lead to unauthorized access or exposure of local file metadata and session content. - [PROMPT_INJECTION]: The skill processes external, untrusted user materials to generate text, which represents an indirect prompt injection surface.
- Ingestion points: User-provided source materials for generating prefaces and reports, processed as described in
SKILL.md. - Boundary markers: Not present; there are no specific instructions or delimiters used to prevent the agent from following commands embedded within user-provided text.
- Capability inventory: Shell command execution and file system read/write via bundled Python scripts.
- Sanitization: None detected; the skill interpolates user input directly into generation prompts without validation.
Audit Metadata