The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill clones a repository from an untrusted GitHub user account seleman66eeddwegger3-art and downloads Python scripts and system configurations via curl from a local network HTTP server.
[COMMAND_EXECUTION]: The instructions involve executing system-level commands like launchctl and systemctl to manage background processes, and using sed to modify sensitive system configuration files.
[PRIVILEGE_ESCALATION]: The skill requires sudo access to modify the system's Redis configuration (/opt/homebrew/etc/redis.conf) and to enable user lingering (loginctl enable-linger), allowing services to run outside of active user sessions.
[PERSISTENCE]: The skill installs persistent background daemons on both macOS (via LaunchAgents in ~/Library/LaunchAgents/) and Linux (via systemd user services in ~/.config/systemd/user/) that are configured to automatically start at load and restart on failure.
[DATA_EXFILTRATION]: The instructions explicitly direct the user to disable Redis security protocols by setting protected-mode no and binding to 0.0.0.0. This exposes the message bus—and any sensitive task data or API keys passing through it—to the entire local network without authentication.
[INDIRECT_PROMPT_INJECTION]: The skill creates a high-risk attack surface by ingesting raw JSON data from a network-exposed Redis queue and passing it directly to an LLM API without sanitization or boundary markers.
Ingestion points: Redis inbox:<NODE_NAME> queue processed in worker_node.py.
Boundary markers: Absent; the skill relies on simple JSON parsing.
Capability inventory: Network operations (requests.post), local file writing (async_debate.md), and Redis list operations (r.lpush).
Sanitization: Absent; the skill performs no validation on the messages content before LLM execution.

hermes-agentmesh-async-bus