The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute local Python scripts for periodic tasks, specifically python agents/scripts/weekly_review.py and agents/scripts/stale_contacts.py.
[DATA_EXFILTRATION]: The skill centralizes and accesses highly personal information across various files, such as relationship histories, contact details, and private goals (network/contacts.jsonl, network/interactions.jsonl, operations/goals.yaml). While these files are intended for user organization, they represent a significant data exposure surface.
[PROMPT_INJECTION]: The skill architecture is susceptible to indirect prompt injection through the ingestion of external data. It captures bookmarks, research, and learning materials into the knowledge/ directory and develops content ideas from these sources without explicit sanitization or instruction boundary markers.
Ingestion points: Data entering the system through knowledge/ (bookmarks/research) and network/interactions.jsonl (external interaction logs).
Boundary markers: No specific delimiters or warnings to ignore embedded instructions are provided in the skill documentation.
Capability inventory: The skill utilizes file system read/write operations and shell-based script execution (Python).
Sanitization: There is no evidence of input validation, escaping, or filtering for content ingested from external sources.
[EXTERNAL_DOWNLOADS]: The skill references documentation and resources from the author's GitHub repository and Anthropic's official engineering guide.

digital-brain