The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from external websites, which serves as a significant attack surface for indirect prompt injection.
Ingestion points: The skill utilizes commands such as vibium text, vibium html, vibium map, and vibium a11y-tree to retrieve content from active browser pages into the agent's context (SKILL.md).
Boundary markers: There are no instructions or recommended patterns for using delimiters or boundary markers to isolate web content from agent instructions.
Capability inventory: The skill possesses high-impact capabilities, including vibium click, vibium type, vibium fill, vibium eval (JavaScript execution), and vibium storage (data extraction).
Sanitization: No sanitization, filtering, or escaping mechanisms for the ingested web content are described in the instructions.
[COMMAND_EXECUTION]: The skill facilitates the execution of the vibium CLI tool and provides logic for resolving the binary path from local or global locations.
[DYNAMIC_EXECUTION]: The vibium eval command allows the agent to execute arbitrary JavaScript code within the browser context to interact with the DOM or extract data.
[DATA_EXFILTRATION]: The vibium storage command allows for the export of sensitive browser session data, including cookies, localStorage, and sessionStorage, to a local JSON file.
[EXTERNAL_DOWNLOADS]: The skill documentation suggests installing the vibium package via npm install -g vibium. Additionally, it supports connecting to remote browsers via WebSocket URLs (e.g., ws://remote-host:9515/session).

browser