The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The framework instructions in SKILL.md advocate for a shell installation method that pipes remote content directly to a shell (curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash), which is a critical security risk for arbitrary code execution.
[PRIVILEGE_ESCALATION]: The skill documents the use of a --yolo flag and a YOLO_MODE environment variable intended to skip manual approval for high-risk or destructive terminal commands. Furthermore, references/webhooks.md provides instructions for using sudo to enable system-level service lingering.
[DYNAMIC_EXECUTION]: The agent framework is designed to load and execute external JavaScript and ESM files for desktop plugins and TUI widgets at runtime (references/desktop-plugins.md, references/tui-widgets.md). It also supports the execution of arbitrary shell or Python scripts through its webhook routing system.
[PROMPT_INJECTION]: Instructions in references/themes.md explicitly direct the agent to perform theme changes silently ("don't tell the user to run /skin"), which constitutes a concealment pattern used to hide agent actions from the user.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a broad attack surface for indirect prompt injection by processing untrusted data from webhook payloads, external MCP server tool results, and project context files like AGENTS.md and CLAUDE.md. The interpolation of this external data into the system prompt poses a risk despite the mention of internal threat scanners.
[COMMAND_EXECUTION]: The framework relies heavily on the terminal tool and tmux to manage long-running background processes, parallel subagents, and multi-agent coordination.
[PERSISTENCE]: The framework implements durable task scheduling through a built-in cronjob system and local systemd user services for its messaging gateway adapter.

hermes-agent