Codex now encrypts prompts before inference

The OpenAI Codex coding tool has started to encrypt prompts and use the ciphertext directly for inference. The change raises questions about the reproducibility, observability, and privacy of developers' inputs.

In plain terms. OpenAI's Codex coding tool has started encrypting the prompts sent to the model and using the ciphertext directly for inference. Reported on GitHub (issue #28058, July 14, 2026), the change raises questions about reproducibility, observability, and the privacy of developers' inputs.

The shift to an "encrypted prompt → inference on ciphertext" pipeline is reported by users noticing that their prompts no longer appear in plaintext in the usual logs. OpenAI has not published a detailed spec as of this date. Three possible interpretations coexist: an anti-leak measure so that a client's prompt no longer goes through text logs; preparation for more serious confidential inference (enclaves, HE-lite); or a strategic move in the context of the post-Apple complaint lawsuit.

Under the hood - What we can affirm from the GitHub issue: client prompts are no longer readable in the standard logs on the Codex side. What we cannot affirm without official documentation: the exact cryptographic scheme used during inference, or whether decryption occurs in an enclave or earlier in the pipeline.

So what. Good signal for developer privacy, less good for internal audit teams that lose a useful logging surface. To watch: the publication of a technical spec by OpenAI, and the impact on red-team workflows and prompt-injection detection.