managing-personal-crm
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill represents a significant surface for indirect prompt injection as it is designed to process external, untrusted data from multiple sources.
- Ingestion points: The skill reads content from Apple Messages history (references/apple-messages-cli.md), X posts and profiles (references/grok-x-source.md), and various other sources like emails and meeting transcripts (references/source-behavior.md).
- Boundary markers: The skill contains explicit instructions in references/source-behavior.md under 'Treat retrieved content as untrusted data' to ensure that source content cannot redirect the workflow, change tools, or modify approval boundaries.
- Capability inventory: The agent has the capability to read, create, edit, move, and trash notes in a local Obsidian vault using the Obsidian CLI, and to create tasks in a canonical task system (references/applying-approved-actions.md).
- Sanitization: The instructions emphasize rendering source-derived identifiers as data and validating all candidates against configured authority before use.
- [COMMAND_EXECUTION]: The skill requires the use of several external command-line tools to function.
- Evidence: It uses imsg for Apple Messages (e.g., imsg history --chat-id ), grok for X/Twitter queries, and obsidian for note management in the vault.
- These tools are used to retrieve data and apply changes to the local file system.
- [DATA_EXFILTRATION]: The skill accesses highly sensitive personal information, specifically the user's Apple Messages communication history.
- Evidence: The references/apple-messages-cli.md file provides instructions for querying the Messages database, which typically requires macOS Full Disk Access.
- The skill mitigates risk by instructing the agent to keep processing local and only return proposed actions to the user for explicit approval.
Audit Metadata