The Agent Skills Directory

[DATA_EXFILTRATION]: The function upload_image_from_file in 6-text-and-images/SKILL.md allows the agent to read arbitrary local files provided by a path and upload them to Miro's external servers.
Evidence: The implementation uses open(file_path, "rb") followed by a requests.post to https://api.miro.com/v2/boards/{board_id}/images. If an agent is manipulated to access sensitive files like ~/.ssh/id_rsa or .env, this functionality serves as an exfiltration vector.
[REMOTE_CODE_EXECUTION]: The GitHub Actions integration pattern in github-actions-integration/SKILL.md uses unsafe string interpolation of untrusted external data into an executed script.
Evidence: The workflow interpolates ${{ github.event.issue.title }} directly into a Python HEREDOC: issue_title = "${{ github.event.issue.title }}". An attacker can craft an issue title containing characters that break the Python string literal (e.g., "; import os; os.system(...)#) to execute arbitrary commands within the GitHub Actions runner environment.

miro-api