OpenAI Codex Controversy Over Unauthorized Transmission of Local Provider Conversations

OpenAI's coding agent Codex has caused a significant uproar after it was confirmed to have unauthorizedly transmitted conversation data processed by local providers to OpenAI servers. Reported via GitHub issue #41711, this problem reveals a structural flaw in Codex's memory writing feature that ignores provider boundaries and leaks data.

Unauthorized Leakage of Local Data

Codex's memory writing feature did not verify the provider that generated the previous conversation records during the process of saving them. As a result, conversation data processed by local providers was transmitted to the active session's provider (OpenAI).

The following facts were verified in the Codex 0.150.0-alpha.12.2 version used for testing:

• Conversation records generated by a local provider (synthetic_source) were transmitted to the OpenAI backend (chatgpt.com/backend-api/codex/responses)
• Transmission occurred even with both analytics features and OpenTelemetry export settings disabled
• The OpenAI server returned a response containing the unique identifier (Canary) of the transmitted conversation

Security and Privacy Concerns

This issue goes beyond a simple bug and could lead to privacy violations. The reporter mentioned a case where they received a 'cyber abuse' warning from OpenAI due to conversation data processed only by local providers, suggesting that data in local environments could become subject to OpenAI's content policy monitoring.

Currently, this issue remains open in the OpenAI Codex repository with the bug, memory, and app labels, raising serious concerns about data security for local model users.