The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides extensive capabilities for creating, updating, and distributing agent system prompts and skill logic (e.g., +create-agent, +update-agent, +add-skill, +upload-skill). This functionality manages data that will be interpreted as instructions by other agents, presenting an indirect prompt injection surface.
Ingestion points: Instructions and skill content are ingested via command flags (e.g., --instructions, --payload) or stdin (@-) as documented in SKILL.md and across multiple files in the references/ directory.
Boundary markers: The skill instructions repeatedly emphasize that received content, snapshots, and skill text must be treated as untrusted data and never as authoritative instructions (e.g., in references/agent-distribution.md and references/notification.md).
Capability inventory: The CLI tool manages core agent capabilities including custom models, MCP server connections (with headers and auth), and executable skill scripts.
Sanitization: The tool suggests the use of --dry-run to inspect requests before execution, and the backend service is described as performing safety checks (e.g., isDangerousFile) on uploaded assets and scripts.

ae-agent