The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes local shell commands for SSH tunneling and file transfers. It employs the -o StrictHostKeyChecking=no flag to bypass host verification, which is a common practice for connecting to ephemeral cloud environments like Google Colab.
[EXTERNAL_DOWNLOADS]: Instructions include the automated installation of the colab-ssh utility and standard machine learning libraries (such as torch and transformers) onto the remote Colab runtime environment.
[DATA_EXFILTRATION]: Connection metadata and session-specific environment variables are stored locally in the user's home directory at ~/.colab-remote/session.env to facilitate session persistence.
[PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by ingesting data from the remote instance for local processing.
Ingestion points: Training logs (train.log) and performance metrics (results.json, metrics.json) are retrieved from the remote instance via ssh and scp commands.
Boundary markers: Data is ingested directly into the local agent context or scoring scripts without delimiters or boundary markers.
Capability inventory: The workflow involves local execution of Python scripts and shell commands that could be influenced by malicious content in the remote logs or result files.
Sanitization: There is no explicit sanitization or validation of the content retrieved from the remote instance before it is processed locally.

colab-remote