The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses and processes highly sensitive personal data from a local Obsidian vault at '~~/Library/Mobile Documents/iCloud~~md~obsidian/Documents/idanbeck/Personal/Family Food Profile.md'. This file contains health information (allergies), family member details, and preferences.
[DATA_EXFILTRATION]: The skill hardcodes a specific residential address ('2267 Shibley Ave, San Jose') and geographic coordinates, which are passed to external CLI commands.
[COMMAND_EXECUTION]: The skill executes a local binary 'dd-cli' to interact with DoorDash. This includes financial operations such as 'order submit' which charges the default payment method on file.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect injection via data ingested from external DoorDash menu items or store descriptions. Malicious content in these external sources could influence the agent's behavior during the composition of orders.
Ingestion points: Reads external store data and menu items via dd-cli search and dd-cli menu (SKILL.md).
Boundary markers: Instructions specify reading from the structuredContent JSON field to separate data from terminal/UI noise.
Capability inventory: Subprocess calls to dd-cli for cart manipulation and order submission; execution of doordash_skill.py for validation.
Sanitization: The doordash_skill.py script validates food-related logic (allergies, balance) but does not sanitize menu item names or descriptions for malicious prompt content.

doordash