The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Automated scans confirmed the presence of malicious URLs within the reference documents of the provided examples. Specifically, 'https://ijels.com/upload_document/issue_files/70IJELS-1042024-TheArt.pdf' is flagged as a botnet resource and 'blog.auspiciousassociates.com' as a phishing site. These malicious resources are included in the research material used by the distilled personas.
[REMOTE_CODE_EXECUTION]: The shell script 'examples/mrbeast-perspective/scripts/fetch_youtube_subtitles.sh' contains logic to silently install a Python package ('yt-dlp') using 'pip install -q'. Executing shell scripts that install third-party dependencies at runtime is a dangerous pattern that can be exploited for supply chain attacks.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting massive amounts of untrusted data from the web (search results, articles, transcripts) and processing it. This creates a significant surface for indirect prompt injection, as there are no explicit sanitization mechanisms or instructions for the agent to ignore embedded commands in the source material.
[COMMAND_EXECUTION]: The core skill logic uses 'Bash' to execute several local Python and Shell scripts. Additionally, the 'x-mastery-mentor' example leverages 'computer-use' for browser automation. This high-privilege access, when combined with the processing of untrusted external content, presents a severe security risk.
[DATA_EXFILTRATION]: The browser automation capabilities featured in the examples could be manipulated via malicious content to exfiltrate user session tokens or sensitive data visible in the browser window.
[OBFUSCATION]: Automated scanners identified multiple files as malicious, including the primary 'SKILL.md'. An AV scan specifically detected a 'HttpRequest-inf' infection in the file 'examples/trump-perspective/references/research/03-expression-dna.md', indicating the presence of malicious code or network behaviors.

huashu-nuwa