The Agent Skills Directory

[COMMAND_EXECUTION]: The skill launches headless Chrome using the --no-sandbox flag in scripts/cap.mjs. This disables the browser's primary security isolation mechanism, making the host system more vulnerable to exploitation if the browser processes malicious web content.
[COMMAND_EXECUTION]: The scripts/cap.mjs script performs file writes using the target.name parameter without sanitization for directory traversal sequences. This could potentially allow an attacker to write files to unintended locations on the filesystem.
[EXTERNAL_DOWNLOADS]: The bootstrap helper scripts/setup-workdir.sh performs runtime installation of the chrome-launcher and chrome-remote-interface Node.js packages using bun add.
[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection through the rendering of arbitrary URLs. Ingestion points: The script scripts/list-stories.mjs fetches and parses index.json from Storybook servers, and scripts/cap.mjs renders HTML and JavaScript from provided URLs. Boundary markers: None are present to distinguish untrusted content from agent instructions. Capability inventory: The skill possesses file system write capabilities (writeFileSync), network access via Node.js http and https modules, and shell command execution. Sanitization: No validation or sanitization is performed on the ingested story data, URLs, or generated filenames.

agh-ui-screenshot