OpenAI hides Codex agent instructions behind encryption, leaving developers in the dark

What happened

Since early June 2026, OpenAI's Codex has encrypted instructions passed between the orchestrating agent and its subagents via a new 'multi-agent v2' protocol. The encryption is mandatory for GPT-5.6 Sol and Terra variants; the smaller Luna model still provides readable descriptions. The Register reported this on July 15, 2026.

Context and impact

Developers building production Codex agent pipelines have lost the ability to audit internal task delegation. OpenAI has given no official explanation; the community speculates about either preventing rival model distillation or basic user privacy. Several teams have called for more transparency.

Details

• Protocol 'multi-agent v2' encrypts the message argument returned by the model
• Codex forwards only ciphertext; Responses API decrypts it internally for the recipient model
• Sol and Terra (larger GPT-5.6 variants): mandatory encryption
• Luna (smaller variant): instructions remain readable
• OpenAI has not formally documented the protocol
• Developer community calling for preserved auditability