The Agent Skills Directory

[COMMAND_EXECUTION]: The function capture_and_analyze in SKILL.md uses socket.socket(socket.AF_PACKET, socket.SOCK_RAW, ...) to capture live network packets. On Linux and most Unix-like systems, this operation requires root privileges or the CAP_NET_RAW capability, constituting a significant privilege requirement.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting and decoding external network data. 1. Ingestion points: Untrusted packet data is received via sock.recvfrom in SKILL.md. 2. Boundary markers: The skill lacks any delimiters or warnings to the agent to ignore instructions embedded within the captured packets. 3. Capability inventory: The skill is designed to parse headers and display payload previews from Ethernet, IP, and TCP layers. 4. Sanitization: Payload content is decoded using utf-8 with replace errors but is otherwise unsanitized, meaning malicious instructions contained in network traffic could be interpreted as commands by the agent.

stevens-network-protocols