run-parallel-agents-feature-build
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external plan files (e.g., from
.ulpi/plans/) and user requests to generate prompts for specialized sub-agents. If these external plans are compromised, they could inject malicious instructions into the sub-agents' context.\n - Ingestion points: Resolved in Step 0 from
arguments.requestand local project files (e.g.,.ulpi/plans/*.json).\n - Boundary markers: No explicit delimiters or warnings are described for isolating plan content within sub-agent briefs in Step 3.\n
- Capability inventory: Spawning sub-agents via the
Agenttool, which have access toBashand file system writes (facilitated by isolation: 'worktree').\n - Sanitization: The skill lacks explicit sanitization or validation of the plan data before it is used to construct sub-agent instructions.\n- [DYNAMIC_EXECUTION]: The skill dynamically assembles instruction briefs for the
Agenttool based on data extracted from external files at runtime. While this is a core functional feature, it creates a mechanism where instruction flow is determined by external, potentially untrusted content.\n- [COMMAND_EXECUTION]: The skill is configured withBashin itsallowed-tools, permitting the execution of shell commands during the orchestration and validation phases.
Audit Metadata