The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates a self-contained HTML file (/tmp/matrix.html) that includes JavaScript for sorting and filtering logic. It then uses Chrome DevTools (mcp__chrome-devtools__new_page, navigate_page) to execute and render this generated file. This pattern involves the runtime generation and execution of code.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. Untrusted data is ingested from external websites via WebSearch and interpolated directly into the DATA section of the JavaScript template in TEMPLATE.md without explicit sanitization or escaping instructions. A malicious actor could embed payloads in search results (e.g., in product reviews or documentation) that, if incorporated into the matrix's verdict or dimension cells, could attempt to influence the agent's behavior or execute scripts in the rendering browser.
Ingestion points: Untrusted content from external websites via WebSearch results.
Boundary markers: The skill provides structural instructions but lacks explicit boundary markers or instructions to sanitize/escape data interpolated into the HTML/JS template.
Capability inventory: Capabilities include file writing (Write), network access (WebSearch), and browser automation/rendering (chrome-devtools tools).
Sanitization: No sanitization or HTML-escaping logic is implemented for the external research data placed into the HTML template.
[COMMAND_EXECUTION]: The skill utilizes browser automation tools to interact with and render locally generated files. While these tools are listed in allowed-tools, their use to render files containing untrusted external data increases the risk profile of the skill's execution environment.

competitive-matrix