The Agent Skills Directory

[DATA_EXFILTRATION]: The skill requests access to sensitive user credentials, specifically the KAGGLE_API_TOKEN environment variable and the ~/.kaggle/access_token file, to authenticate interactions with the Kaggle MCP server. This is consistent with the vendor's own authentication requirements.
[COMMAND_EXECUTION]: Instructions suggest using CLI tools such as Playwright and yt-dlp to process external URLs and media provided in submissions. Executing these tools on data from untrusted sources creates an attack surface for potential environment exploitation.
[PROMPT_INJECTION]: The skill processes untrusted user-generated content (writeups and artifacts), representing an Indirect Prompt Injection risk.
Ingestion points: Submissions retrieved via list_hackathon_write_ups and external links processed by Playwright/yt-dlp.
Boundary markers: The instructions do not specify explicit delimiters or "ignore previous instructions" guards for the ingested content.
Capability inventory: The skill uses subprocess calls to external CLI tools and performs logic-based ranking and grading.
Sanitization: No specific sanitization or filtering logic is provided within the skill instructions.
[EXTERNAL_DOWNLOADS]: The skill references an external GitHub repository (hamelsmu/evals-skills) as an auditing toolkit and points to a web dashboard hosted on an external Google Cloud Run URL.

hackathon-judging