OpenAI 加密 Codex 代理間的指令——開發者如何偵錯他們看不到的內容？

openai has implemented mandatory encryption for instructions exchanged between codex agents, rendering the internal delegation process unreadable to developers. initially affecting gpt-5.5 before a partial rollback, the restriction now applies strictly to larger gpt-5.6 models sol and terra, leaving users unable to verify how tasks are distributed. reports indicate that the encrypted strings occasionally cause decryption errors and task handoff failures. while unconfirmed by openai, the move likely aims to prevent rival models from distilling valuable agent-to-agent training data, trading user transparency for intellectual property protection.