OpenAI Encrypts Codex Agent Instructions, Raising Developer Auditability Concerns

Summary

OpenAI has begun encrypting multi-agent v2 message payloads in its Codex command-line interface, preventing developers from viewing plaintext instructions passed between parent and child agents. Previously, these instructions were visible in logs and history, allowing developers to debug and audit agent behavior. The change encrypts agent instructions at the API level, with only Codex forwarding ciphertext and individual agents decrypting messages internally.

Developers have raised concerns that the encryption removes critical observability and auditability capabilities. Ignat Remizov, CTO at payment service Zolvat, noted that the encrypted delivery path removes "human-readable task/message text from local rollout history, trace reduction, and parent-side audit/debug surfaces." Without access to this information, developers and maintainers will have less visibility into what instructions agents received and what actions they took.

OpenAI has not publicly explained the motivation for the change, which was merged in a pull request that stated "Why" but provided no reasoning. While privacy hardening and protection against model distillation are plausible motivations, the lack of transparency has fueled speculation that OpenAI is protecting proprietary implementation details of its multi-agent orchestration system.

• Industry concerns focus on losing the ability to verify and understand agent decision-making in production systems
• Some developers speculate the encryption is meant to conceal how OpenAI's multi-agent implementation works from competitors