The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a bash helper script (scripts/manus.sh) that executes several system commands to facilitate API interaction.
It uses curl to make requests to the Manus API at api.manus.im and to download task results.
It employs jq for robust parsing of JSON data from API responses.
It utilizes utility commands such as mkdir, tr, and file for directory management and metadata extraction during file handling.
[EXTERNAL_DOWNLOADS]: The skill is designed to download generated deliverables (e.g., PDFs, CSVs, code) from the Manus content delivery network.
Files are retrieved from private-us-east-1.manuscdn.com and saved locally using curl with a sanitization step for filenames to prevent directory traversal or special character exploitation.
These downloads are integral to the service functionality and originate from official infrastructure.
[DATA_EXFILTRATION]: The skill facilitates the transfer of local files to the Manus service for task processing.
This is achieved by requesting a presigned S3 upload URL from the Manus API and then using curl to PUT the file content to Amazon S3.
This behavior is user-directed and necessary for the 'upload' and 'attachment' features of the agent integration.
[PROMPT_INJECTION]: The skill possesses an inherent surface for indirect prompt injection because it processes output from an external autonomous agent that interacts with the public web.
Ingestion points: Data enters the skill context through the retrieval of task results and deliverables via the get and download commands in scripts/manus.sh.
Boundary markers: The instructions and templates do not specify delimiters or explicit instructions to ignore potentially malicious content embedded within the task results received from the Manus agent.
Capability inventory: The agent environment allows for file system writes, network requests via curl, and execution of the included bash script.
Sanitization: While the skill sanitizes filenames during download, it does not sanitize the content of the retrieved task results before they are presented or potentially used in multi-turn conversations.

manus