OpenAI Encrypts Internal Codex Prompts, Breaking Workflows
OpenAI is now scrambling subagent instructions in Codex logs, shielding internal prompt logic while occasionally preventing its own models from decrypting their tasks.
An update to OpenAI's Codex coding assistant has turned previously visible inter-agent coordination into unreadable gibberish. According to an issue report on the project's GitHub repository, the plain-text instructions the primary AI agent passes to subagents now appear as encrypted strings in session logs.
The mandatory encryption applies to larger models in the GPT-5.6 lineup, specifically Sol and Terra. Meanwhile, the lightweight Luna variant continues to pass commands in the clear. Readable logs were also reportedly restored to the older GPT-5.5 branch after previously being encrypted.
The shift solves a corporate concern—protecting system prompts and intermediate reasoning from scrapers, competitors, or model distillation—while creating practical headaches for users. Developers complaining on GitHub point out that obscuring these payloads destroys workflow visibility. More critically, the encryption routine occasionally breaks execution entirely: subagents sometimes fail to decrypt the incoming instructions, throwing errors even when the main agent and subagent share the exact same model tier.
OpenAI has not publicly stated the reasoning behind the change. Observers suspect it serves a dual purpose of proprietary prompt defense and server-side privacy, aligning with the company's existing practice of passing encrypted intermediate states between API calls.
Source: habr
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.