The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation instructs the agent to download and execute an installation script directly from a remote URL using a piped shell command (curl -fsSL https://deeplake.ai/install.sh | bash). This practice allows for arbitrary code execution from an external source without verification of the script's integrity by a package manager.
[COMMAND_EXECUTION]: The skill advises the use of sudo apt-get install ffmpeg for installing system dependencies. Directing an agent to use sudo for administrative tasks introduces a risk of privilege escalation and potential system compromise.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading and ingesting data from external platforms such as HuggingFace (_huggingface key) and other public dataset repositories, which may contain untrusted content.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its core functionality of ingesting and processing content from external files and datasets.
Ingestion points: The client.ingest() method processes local files (videos, PDFs, images) and remote datasets (HuggingFace).
Boundary markers: There are no explicit instructions or delimiters provided to ensure the agent ignores commands or instructions embedded within the ingested data.
Capability inventory: The skill is granted powerful tools including Bash, Write, Edit, and WebFetch which could be exploited if malicious instructions are successfully injected via the data pipeline.
Sanitization: No validation or sanitization of the input data is documented before the content is passed into the processing pipeline.

deeplake