AI TechnologyOpenAIJul 15, 2026 15:21 UTC

OpenAI Codex Encrypts Instructions Between Agents

OpenAI's coding tool 'Codex' changed its specifications from early June 2025 to encrypt the content of instructions passed from main agents to sub-agents. As a result, developers can no longer monitor the task delegation process between AI agents from the outside. For large-scale models 'Sol' and 'Terra', this encryption is mandatory.

OpenAI Codex Encrypts Instructions Between Agents

OpenAI's coding tool 'Codex' changed its specifications from early June 2025 to encrypt the content of instructions exchanged between AI agents. As a result, developers can no longer track what tasks main agents are assigning to sub-agents from the outside.

Originally, Codex adopted a 'multi-agent' configuration where multiple AI agents collaborate to handle complex programming tasks. In this structure, the main agent as the parent role breaks down work into smaller pieces and distributes them to sub-agents as child roles. Previously, developers could monitor this delegation process, but the current change has effectively blocked that pathway.

As confirmed in this case, for Codex's larger models 'Sol' and 'Terra', this encryption is not optional but mandatory. In other words, developers using these models essentially have no means to reference the contents of internal instructions.

The significance of this change becomes clear when viewed from the perspective of the developer's working environment. When building and debugging AI systems, the ability to visualize the flow of processing directly impacts quality assurance and early problem detection. With agent-to-agent instructions no longer visible, identifying the cause of unintended behavior becomes more difficult.

On the other hand, this design also has another dimension. Making agent-to-agent communication unreadable from outside can be interpreted as a security intention to protect the system's internal logic. However, at this point, there is no confirmed information that OpenAI has officially explained the detailed intention or technical rationale behind this change.

As the use of multi-agent systems expands, how to balance transparency and security in design is increasingly positioned as an important issue in AI development. This change can be seen as one example of raising such a question to the developer community. Going forward, whether OpenAI provides an explanation for this design decision will be a key point to watch.

#OpenAI#AIAgent#Codex#MultiAgent#GenerativeAI#AIDevelopment#Security
AI issue Staff

This article is an original work independently written and edited by the AI issue editorial team based on factual reporting. © AI issue. Unauthorized reproduction, redistribution, or use for AI training is prohibited.

Comments

Log in to comment