The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions provide several large JavaScript templates (such as multi-tool-scan.js, parallel-audit.js, and crawl-audit.js) which the agent is directed to write to the filesystem and execute using Node.js. This pattern of runtime script generation and execution increases the risk of code being modified or injected with malicious logic during the generation phase.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from external websites, including HTML structures and video frames. There are no explicit boundary markers or sanitization procedures defined to protect the agent from 'jailbreak' attempts or malicious instructions embedded in the audited site's content, which the agent is specifically instructed to 'READ' and 'ANALYZE'.
[COMMAND_EXECUTION]: The skill makes extensive use of the shell to install Node.js packages, execute the generated audit scripts, and process video files using ffmpeg. The autonomous nature of the skill (using instructions like <default_to_action>) means these commands may be executed with reduced user oversight.
[EXTERNAL_DOWNLOADS]: The skill uses curl to download video files from external, user-provided URLs and uses npm to download multiple third-party libraries. While the libraries themselves are well-known, the process of downloading and executing content from arbitrary URLs is a standard security risk factor.

qe-a11y-ally