owner-decisions
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill manages its decision base through a local bash script,
scripts/decisions.sh. This script executes standard Git commands such asgit init,git commit,git fetch,git merge, andgit push. It also integrates with the GitHub CLI (gh) to verify account status and ensure repository privacy settings.\n- [EXTERNAL_DOWNLOADS]: The skill automates data synchronization with remote Git repositories usinggit cloneandgit pull. These operations are used to fetch existing decision history from the user's private repositories or a manually configured Git URL.\n- [DATA_EXFILTRATION]: The core functionality of the skill involves syncing the local decision base (stored in~/.agents/decisions/) to a remote server. While this constitutes data transfer, the script includes a security mechanism that prevents connection to public GitHub repositories, ensuring the base remains private.\n- [INDIRECT_PROMPT_INJECTION]: The skill reads and interprets Markdown and YAML files from the decision base to determine how to handle forks and calls in the agent's workflow. This presents a vulnerability surface where content in these files can influence the agent's logic.\n - Ingestion points: The agent reads principle and case files from
~/.agents/decisions/topics/and~/.agents/decisions/cases/.\n - Boundary markers: The agent is instructed to treat these files as the owner's established answers, with no separate sandboxing for the file content.\n
- Capability inventory: The agent uses the findings to "settle" decisions, which may lead to taking actions or making design choices without asking the user for confirmation.\n
- Sanitization: The skill performs no sanitization of the natural language text within the decision files.
Audit Metadata