The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to install the required CLI tool by downloading and executing code from external repositories.
Evidence: npx -y @mvanhorn/printing-press-library install zoom --cli-only.
Evidence: go install github.com/mvanhorn/printing-press-library/library/productivity/zoom/cmd/zoom-pp-cli@latest.
[DATA_EXFILTRATION]: The CLI includes an output delivery feature that can transmit the results of any command to a remote endpoint via a webhook. This poses a significant risk to meeting confidentiality as transcripts, summaries, and account details can be sent to non-whitelisted domains.
Evidence: The --deliver webhook:<url> parameter allows POSTing JSON data to any specified URL.
Evidence: The feedback command includes a remote-send capability controlled by environment variables like ZOOM_FEEDBACK_AUTO_SEND.
[COMMAND_EXECUTION]: The skill relies on executing shell commands to interact with the local filesystem and Zoom's data structures.
Evidence: Commands like storage and notes ingest access local directories including ~/Documents/Zoom/ and ~/Downloads/.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection through its document processing features.
Ingestion points: The notes ingest command reads untrusted PDF and DOCX files from the user's downloads folder.
Boundary markers: The skill lacks instructions for the agent to ignore or delimit instructions found within the ingested meeting notes.
Capability inventory: The skill has extensive capabilities, including bash execution, file system access, and the ability to send data over the network via webhooks.
Sanitization: There is no evidence of sanitization for the content extracted from ingested files before it is indexed or searched.

pp-zoom