Codex Logging Bug Caused Up To 640 TB/Year Of SSD Writes, Mostly Fixed
The bug exposed a real-world risk of rapid SSD wear from AI agent logging, and the fix shows OpenAI can respond quickly to infrastructure-level issues reported by the user community.
Reporting from 1 sources: GIGAZINE.
OpenAI's Codex AI coding assistant had a logging bug that could write approximately 640 TB of data per year to a local SSD, potentially exceeding the endurance of a typical 1 TB drive in under a year. The issue was reported by GitHub user 1996fanrui, who observed about 37 TB written over 21 days of use, with Codex SQLite logs as the primary source. Investigation showed the database retained only about 680,000 logs, but the SQLite row ID exceeded 5.5 billion, indicating repeated cycles of inserting large amounts of logs and immediately deleting them. Most writes came from TRACE level detailed logs recording function calls and variable values. OpenAI released version 81150 of win-codex a few days after the report. Thibault Soti, OpenAI's head of core products, said on X: "There was a bug in Codex. Fixed it." GitHub user wfy-op reported that before the update there were 28,000 to 46,000 writes in 30 seconds, dropping to about 6,700 after the update. OpenAI claims the latest version reduces write volume by about 85%. However, TRACE level logs are still saved to the database, and wfy-op noted write volume, while reduced, is still not normal low-frequency log behavior.
The bug was reported in a GitHub issue on the openai/codex repository. The reporter, 1996fanrui, measured approximately 36,211 new log rows written in a 15-second window, with rows being inserted, indexed, written to the WAL, and then deleted continuously while the retention count stayed unchanged. The analysis traced the majority of writes to TRACE level logs that record function calls and variable values, along with a large volume of internal file monitoring logs.
The report suggested several changes: not persisting TRACE logs by default, setting database size limits, and saving only the minimum necessary data instead of full WebSocket payloads. OpenAI released version 81150 of win-codex a few days later. Thibault Soti, OpenAI's head of core products overseeing both ChatGPT and Codex, said on X: "There was a bug in Codex. Fixed it."
GitHub user wfy-op reported that before the update there were 28,000 to 46,000 writes in 30 seconds, dropping to about 6,700 after the update. OpenAI claims the latest version reduces write volume by about 85%. However, TRACE level logs are still saved to the database, and wfy-op noted write volume, while reduced, is still not normal low-frequency log behavior.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.