gpt-5-6-best-practice
Pass
Audited by Gen Agent Trust Hub on Jul 14, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [SAFE]: The skill implements a routing and evaluation overlay for optimizing task execution cost without bypassing safety filters. It contains no instructions that attempt to override agent behavior or bypass safety guidelines.
- [EXTERNAL_DOWNLOADS]: The skill references technical documentation and benchmarks from OpenAI and recognized industry analytics platforms such as Artificial Analysis, CodeRabbit, and Qodo to inform its routing logic. These are documented neutrally as sources of truth for model performance characteristics.
- [COMMAND_EXECUTION]: The framework utilizes named agent profiles in local configuration files under
.codex/agents/to manage sub-task delegation. This orchestration pattern uses specific sandbox settings to limit capabilities, such as creating read-only workers for exploration and write-enabled workers for bounded implementation. - [PROMPT_INJECTION]: The instructions explicitly reinforce the authority of system-level safety guidelines and user-specified constraints, stating they remain authoritative over the skill's own logic.
- [DATA_EXFILTRATION]: The skill promotes data minimization by instructing the agent to return distilled evidence rather than raw logs or unnecessary repository content between tasks, reducing the risk of unintentional data exposure.
Audit Metadata