OpenAI's Codex CLI encrypts sub-agent messages by default, drawing a compliance backlash

A default change in OpenAI’s Codex CLI that encrypts messages between its sub-agents has resurfaced as an enterprise compliance complaint, because it leaves local audit logs unreadable. The company’s command-line coding agent now stores only encrypted content for inter-agent messages, so the plaintext that once appeared in session logs is gone unless OpenAI decrypts it.

What changed since the feature first drew notice: the debate has moved from a quiet code merge to open, growing pushback from developers who say the design breaks compliance requirements. The encryption shipped in a pull request merged June 5, 2026, covering the experimental MultiAgentV2 feature when it runs on OpenAI’s Sol or Terra models. It encrypts the spawn_agent, send_message and followup_task calls and leaves the plaintext content field empty.

For regulated buyers, that is the problem. Teams in finance, healthcare and government are often required to keep readable local records of what automated systems do; an audit trail they cannot decrypt fails that test. A GitHub issue opened June 13, 2026 documents the regression and proposes a fix, deliver encrypted content to the model but write a separate plaintext copy for audit, and includes a prototype patch. The issue remains open.

OpenAI has not shipped the dual-field fix, and it has not publicly explained the security rationale for encrypting the messages by default. Whether the company adds an audit-friendly plaintext option, or holds the line and pushes compliance-sensitive teams toward other agent frameworks, is what enterprise users are now watching.