The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/mp-downloader.py employs importlib.import_module to load downloader client classes at runtime. While the specific module paths are retrieved from a hardcoded internal map (PROVIDER_CLASSES), this constitutes dynamic code loading behavior.
[COMMAND_EXECUTION]: The tasks.add.direct action allows the AI agent to read the contents of local files by specifying a path in the content argument when torrent_file is set to true. This functionality could be exploited to read sensitive system files if the agent is manipulated into accessing unauthorized paths.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external downloader instances, such as torrent names, tracker information, and tags, which are ingested into the agent's context. This creates a potential surface for indirect prompt injection if the external data contains malicious instructions.
Ingestion points: Data is retrieved from the downloader via multiple actions including tasks.list, tasks.trackers, and tasks.peers in scripts/mp-downloader.py.
Boundary markers: The instructions do not define explicit delimiters or warnings to treat downloader-provided strings as untrusted data.
Capability inventory: The skill utilizes execute_command to run a local Python script capable of file system reads and downloader state modification.
Sanitization: The script actively filters known sensitive keys (e.g., passwords, tokens) but does not sanitize the content of descriptive fields that may contain injection payloads.
[SAFE]: The script implements proactive security measures, including the _is_sensitive_key and _jsonable functions, which are designed to identify and remove credentials, API keys, and session tokens from the data returned to the agent.

downloader-operation