The Agent Skills Directory

[PROMPT_INJECTION]: The skill demonstrates a vulnerability surface for indirect prompt injection. It ingests existing data from the target uCoz website—such as page templates, menu configurations, and mail form fields—without using boundary markers or sanitization. If an attacker manages to embed malicious instructions within the site's content, the agent might interpret and execute them while performing its tasks.
Ingestion points: In references/PLAYBOOK.md, the agent is instructed to read homepage content (page_get), template code (read_template), and mail form configurations (mail_list_fields).
Boundary markers: No delimiters or warnings to ignore embedded instructions are used for the ingested data.
Capability inventory: The skill possesses high-impact capabilities, including modifying templates, managing FTP credentials via ftp_tool, and altering site indexing settings via modules_tool.
Sanitization: No methods for escaping or validating external site content are described before the data is processed.
[COMMAND_EXECUTION]: The skill uses dynamic script generation to create website favicons. It provides a PowerShell template that utilizes the System.Drawing assembly to programmatically generate an image file and save it to the local workspace. While the specific script provided is intended for benign branding purposes, generating and executing scripts at runtime is a pattern associated with more advanced attack vectors.

ucoz-landing-skill