Inside Codex: 42KB of JSON Reveals Hidden Prompts, Tools, and Context
The blog measured a real Codex request for a tiny 16‑character prompt and found a ~43 KB JSON payload, roughly 9,400 tokens, of which only 0.3 % are the user input. The bulk consists of internal instructions, tool definitions, permissions, skill metadata, and repository context, exposing hidden data sent to the model.