OpenAI has encrypted the internal instructions passed between its AI agents, meaning developers can no longer read how tasks are delegated within the system.
Since early June, the coding tool Codex has replaced readable task descriptions with unreadable strings in the session history. This change affects how agentic systems operate. These tools break down work and hand off parts to subagents, making visibility into that process vital for debugging.
A GitHub bug report highlights the issue, asking OpenAI to store a readable copy of the task locally alongside the encrypted version.
Previously, GPT-5.5 did not allow developers to disable encryption via a dedicated toggle. OpenAI has since restored the readable path for that model. The mandatory encryption now applies only to the larger GPT-5.6 variants, named Sol and Terra. The smallest variant, Luna, continues to use the open path.
The new system also appears unreliable. Several developers report failed handoffs because the content cannot be decrypted. This happens even when the main agent and subagent use the same model.
Why OpenAI is hiding agent chatter
OpenAI has not explained the rationale behind this shift. Only the technical change itself has been confirmed.
Community members suspect the company treats these prompts as reasoning traces and wants to prevent rivals from training on them. Zhipu AI’s open GLM-5.2 model was recently suspected of being distilled from GPT-5.5 and Opus 4.8. Agent-to-agent communication serves as valuable training data that can lift a weaker model toward a stronger one’s level. Encrypting it keeps that material out of competitors’ hands.
A simpler reason is equally plausible. OpenAI’s API already encrypts intermediate states so they can be forwarded in follow-up requests without storing plaintext on its servers. Confirmation is still pending on whether the move targets distillation protection, data privacy, or both.
What it means
Developers lose the ability to audit internal delegation logic. Without seeing the handoff instructions, tracking bugs in complex workflows becomes harder. The forced encryption on Sol and Terra adds a layer of opacity that was previously available for inspection.
