The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install an external tool globally using npm install -g claude-flow@alpha in the Quick Start section.
[REMOTE_CODE_EXECUTION]: Setup instructions involve npx claude-flow@alpha mcp start, which downloads and executes remote code from the npm registry at runtime.
[COMMAND_EXECUTION]: Multiple tools like mcp__claude-flow__parallel_execute and mcp__claude-flow__task_orchestrate take a command or task parameter that is directly mapped to shell execution (e.g., command: "npm run test:unit", command: "npm run test:security"). This allows for arbitrary command execution across the swarm.
[PERSISTENCE]: The skill implements automated triggers and persistent coordination rules via mcp__claude-flow__automation_setup and mcp__claude-flow__trigger_setup. These create long-running hooks that respond to events such as file-changed or PR-created (e.g., trigger: "file-changed", action: "run-tests").
[INDIRECT_PROMPT_INJECTION]: The orchestration engine creates a large attack surface where untrusted data (like task descriptions or external search results) can influence shell commands executed by specialized agents.
Ingestion points: task and command parameters in orchestration tools; data parameter in pattern_recognize.
Boundary markers: None present to separate instructions from untrusted data.
Capability inventory: Shell command execution via parallel_execute; filesystem access via memory_backup.
Sanitization: No validation or sanitization of input strings before interpolation into execution contexts is evident.
[DYNAMIC_EXECUTION]: The skill uses mcp__claude-flow__neural_train to dynamically train agent coordination patterns based on training_data, which can modify the swarm's logic at runtime.
[DATA_EXFILTRATION]: The mcp__claude-flow__memory_backup tool allows writing internal agent memory to arbitrary local paths (e.g., /workspaces/claude-code-flow/backups/swarm-memory.json), which could be used to stage data for exfiltration.

swarm-advanced