OpenAI’s Codex now encrypts instructions between AI agents, leaving developers blind to internal delegation

What changed

OpenAI has updated Codex to encrypt the instructions that its main AI agent sends to subordinate agents. This change started in early June and now prevents developers from seeing or tracking how tasks are divided and delegated among the AI’s internal subagents. The encryption is mandatory on larger GPT-5.6 models called Sol and Terra, making this a baseline behavior for those platforms.

Why builders should care

Developers using Codex can no longer inspect the step-by-step delegation of tasks inside the AI. For practical purposes, this removes visibility into how the AI breaks down and distributes work internally. Builders who previously experimented with or debugged multi-agent workflows will lose insight into the logic behind internal task assignments. This reduces transparency and could complicate troubleshooting or refining agent behavior for complex coding or automation projects.

The practical takeaway

The encryption backs OpenAI’s move to treat multi-agent instruction exchange as a black box rather than an open process. This puts more trust and control in OpenAI’s hands at the cost of developer oversight. Anyone relying on internal command visibility to optimize or audit AI task management must now rethink their approach. Debugging multi-agent AI flows is likely to rely more on outputs and less on inspecting internal messaging.

What to watch next

It will be important to track if OpenAI extends this encryption mandate to smaller Codex models or other AI tools beyond GPT-5.6 variants. Also watch how developers and enterprises respond to the loss of internal transparency. Potential pressure could build for enhanced auditability or alternative ways to verify correct task delegation without sacrificing encryption. Meanwhile, software relying on agent orchestration should prepare for a more opaque operating environment.

AI Quick Briefs Editorial Desk