The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core purpose is to facilitate Agent-to-Agent (A2A) communication, which inherently creates a surface for indirect prompt injection attacks. If one agent in the network is compromised or ingests untrusted data, it can propagate malicious instructions to other agents via message routing or shared state synchronization.
Ingestion points: The send-message command (specifically the --message argument) and the distributed shared state management system documented in SKILL.md.
Boundary markers: The documentation does not specify the use of delimiters, XML tags, or instruction isolation techniques to prevent agents from executing instructions contained within message payloads.
Capability inventory: The management script scripts/llmops_manager.py appears to have broad capabilities including agent registration, message routing, and state persistence.
Sanitization: There is no evidence of message sanitization, validation, or filtering to ensure that communicated content does not contain unauthorized instructions.
[COMMAND_EXECUTION]: The usage examples in SKILL.md demonstrate the execution of shell commands to interact with a management script (scripts/llmops_manager.py). This script is responsible for critical functions such as registering agents and distributing messages.
[REMOTE_CODE_EXECUTION]: The skill relies on an external Python script (scripts/llmops_manager.py) which is not included in the provided files. This constitutes an unverifiable dependency that is expected to be present in the execution environment.

a2a-communication-manager