rowboat
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external work data such as emails, meeting transcripts, and Markdown notes, which represents a potential surface for indirect prompt injection.
- Ingestion points: The Rowboat application ingests content from connected integrations (Gmail, meeting providers) and local Markdown source files.
- Boundary markers: The graph builder implementation (referenced in
references/upstream.md) uses aBATCH_SIZE = 1to process files individually, reducing cross-file entity contamination. - Capability inventory: The skill uses
Bash,Read,Grep, andGlobto interact with files and relies on the Rowboat application for data processing. - Sanitization: The instructions explicitly forbid inferring owner identity from email headers and require verification of source-grounded implementation details before making changes.
- [COMMAND_EXECUTION]: The skill includes a Python script
scripts/preflight.pyfor offline file integrity checks. This script is implemented using only standard libraries and employs security best practices, such asos.O_NOFOLLOWand directory file descriptors (dir_fd), to prevent symlink attacks and ensure secure file traversal. It also contains a hardcoded list of sensitive file paths (e.g.,.ssh,.env,credentials) to prevent accidental or malicious reading of local secrets.
Audit Metadata