parallel-goal-workflows
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent orchestration framework that processes user-defined goals. While this creates a data ingestion surface, the skill includes several architectural mitigations.
- Ingestion points: User-provided goals and constraints entered via skill triggers described in
SKILL.md. - Boundary markers: The instructions explicitly require the creation of 'Local Briefs' using structured headers (Local goal, Relevant context, Boundary, Deliverable and evidence, Pause if) to isolate the sub-agent's execution environment.
- Capability inventory: The skill utilizes the host's ability to spawn sub-agents and manage nested goals as documented in
references/codex-nested-subagents.md. - Sanitization: The
SKILL.mdfile mandates that the 'Main Agent' synthesize user intent into a clean task-local brief while specifically excluding raw transcripts, skill body text, and orchestration triggers. - [COMMAND_EXECUTION]: The skill includes a Python script (
evals/check_identity_packet.py) which is a linting tool used to verify that the skill's instructions remain effective and that orchestration context is not leaked. The script performs read-only operations on local skill files (SKILL.md,agents/openai.yaml) and does not interact with sensitive system files or execute external commands. - [DATA_EXFILTRATION]: No evidence of credential theft or unauthorized data transmission was found. External URLs point to the author's official GitHub repositories and legitimate developer documentation for OpenAI Codex.
Audit Metadata