The Agent Skills Directory

COMMAND_EXECUTION (LOW): The skill executes tailscale status --json via subprocess.run in tailscale_localsend.py. While this is a standard administrative command for the intended purpose, it represents local system command execution.
DATA_EXFILTRATION (MEDIUM): The send method in tailscale_localsend.py can read any file path accessible to the agent and transmit its contents via HTTP to a network peer. An agent could be manipulated through prompt injection to send sensitive system files (e.g., SSH keys, configuration files) to a malicious peer discovered on the network.
DATA_EXPOSURE (MEDIUM): The receive method starts an unauthenticated HTTPServer that accepts file uploads from any network peer and writes them to the local disk. This creates a surface for attackers to place malicious files on the host system or perform a denial-of-service attack by filling disk space.
EXTERNAL_DOWNLOADS (LOW): The documentation mentions an integration with epistemic-arbitrage, which is an unknown and potentially unverified external package.
INDIRECT_PROMPT_INJECTION (LOW): The skill ingests untrusted data from network peers via UDP multicast discovery and HTTP file reception. If the agent later parses or follows instructions contained within the metadata or contents of these files, it creates a surface for indirect prompt injection.

tailscale-localsend