The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches JavaScript components from 'cdn.jsdelivr.net' and clones the source code from 'github.com/alibaba/page-agent.git'. These resources originate from well-known and reputable technology providers.
[COMMAND_EXECUTION]: The skill instructions include standard development operations such as npm install, npm ci, and git clone. These commands are necessary for the installation and development workflows described in the skill.
[INDIRECT_PROMPT_INJECTION]: As a browser-based agent that reads and acts upon the Document Object Model (DOM) of web pages, this tool has an inherent attack surface for indirect prompt injection. If a web page contains malicious instructions in its text or metadata, the agent could potentially be manipulated into performing unintended actions. The skill documentation acknowledges this by mentioning allowlisting and data-masking configurations to mitigate such risks.
Sanitization/Evidence Chain for Indirect Prompt Injection:
Ingestion points: Reads the current webpage DOM as text to interpret UI elements.
Boundary markers: Not explicitly defined in the basic instructions, though configuration options for locking down access are referenced.
Capability inventory: Can perform 'click', 'fill', and 'execute' actions against DOM elements.
Sanitization: Mentions the existence of data-masking hooks and allowlists to restrict agent interactions.

page-agent