The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/manage_knowledge.py includes an upload_file function that reads local files using path.read_text() and sends the contents to a remote server via client.upload_knowledge_content. This capability can be used to exfiltrate sensitive files if the agent is directed to upload them to a rogue server URL.\n- [COMMAND_EXECUTION]: The skill documentation instructs the agent to execute a local Python script scripts/manage_knowledge.py using the uv tool. The script processes several command-line arguments, including file paths and URLs, which the agent might populate with untrusted or malicious input.\n- [PROMPT_INJECTION]: The skill implements a workflow for searching and retrieving content from external knowledge bases, which serves as an ingestion point for untrusted data (Indirect Prompt Injection).\n
Ingestion points: Content retrieved from search queries in scripts/manage_knowledge.py and AgentOSClient.search_knowledge.\n
Boundary markers: No delimiters or instructions to ignore embedded commands were found in the script's handling of search output.\n
Capability inventory: The skill allows file system reads, local script execution, and network POST requests.\n
Sanitization: The script does not perform sanitization or validation of the content retrieved from the knowledge base before it is returned to the agent context.

agentos-api-knowledge