Codex Encryption Update Removes Human-Readable Agent Audit Trails

Codex starts encrypting prompts, uses ciphertext for inference instead

A recent update to OpenAI Codex encrypts MultiAgentV2 message payloads, inadvertently stripping away the human-readable audit trail for subagent tasks. While intended to enhance privacy, this change makes it impossible for users to inspect what instructions were delegated to child agents or debug failures after the fact. The community argues that maintaining local auditability is essential to prevent building opaque AI systems that users cannot monitor or understand.

"Guys, we don’t want to build Skynet and then be unable to audit what it’s doing."