roblox-build-orchestrator
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted data from external sources and coordinates multiple agents.
- Ingestion points:
scripts/run_log.pyreads event logs (events.jsonl) and metadata. Workers likeroblox-market-scoutandroblox-asset-sfxingest data from web searches and external APIs (WebFetch,WebSearch). The manager agent reads manifests produced by subagents. - Boundary markers: The skill explicitly enforces a 'write boundary' where subagents are restricted from using Studio tools via
disallowedTools: mcp__robloxstudio-mcp. Interaction with the Roblox DataModel is restricted to a single thread. - Capability inventory: The system performs file operations, localized Bash commands, and web fetching. The primary manager agent has access to Roblox Studio tools (
robloxstudio-mcp). - Sanitization: Data is processed through standard JSON parsing (
json.loads) inrun_log.pyand structured manifests. - [COMMAND_EXECUTION]: The skill provides instructions for setting up agents and uses a Python script for run management.
- Evidence:
SKILL.mdincludes a setup command:mkdir -p ~/.claude/agents && cp ~/.claude/skills/roblox-build-orchestrator/agents/*.md ~/.claude/agents/. This is a standard procedure for registering subagents in the Claude Code environment. - Evidence:
scripts/run_log.pyis a utility script using standard Python libraries for logging and reporting, with no suspicious subprocess or remote execution patterns.
Audit Metadata