The Agent Skills Directory

[Indirect Prompt Injection] The skill implements context.callAgent() to analyze email content. Maliciously crafted emails (e.g., in ai-classifier.ts or urgent-watcher.ts) could contain instructions intended to override the classifier's logic, potentially tricking the agent into marking spam as urgent or bypassing specific filters. This is a common risk with LLM-based data processing.
[Command Execution] The skill provides the capability to write and modify TypeScript files in agent/custom_scripts/listeners/. While intended for creating listeners, this tool could be used to write arbitrary code if the agent is misled by a prompt injection attack.
[Data Exposure Risk] By monitoring and processing all incoming emails, the skill handles sensitive personal and professional data. The templates include logic for reading subjects, senders, and body content to determine automated actions like archiving or starring.

listener-creator