OpenAI reveals ChatGPT bug exposed 1.2% of paid users' billing data

OpenAI has released a detailed post-mortem on the bug that forced ChatGPT offline on March 20, acknowledging that a flaw in an open-source library exposed payment-related information for a small fraction of its Plus subscribers. The root cause was traced to the redis-py client library for Redis Cluster, which OpenAI uses to cache user data. A server change introduced early Monday morning caused a spike in Redis request cancellations, corrupting connections in a way that could return another user’s cached data as valid. While the issue was most visible as some users seeing snippets of other people’s chat history titles, a deeper investigation revealed something more worrying: for a nine-hour window, about 1.2% of active ChatGPT Plus subscribers had their first and last name, email address, payment address, credit card type, last four digits of their card number, and expiration date potentially visible to other users. Full card numbers were never exposed. The company believes the number of users whose data was actually seen by someone else is “extremely low,” requiring them to have opened a misdirected subscription confirmation email or viewed their account management page during that specific period. OpenAI has notified the affected users. The bug itself is a textbook race condition in asynchronous code—when a request is canceled after being pushed to the incoming queue but before its response is popped from the outgoing queue, the connection gets corrupted and the next unlucky request can receive leftover data. In most cases this would just throw an error, but sometimes the garbage data looked valid enough to pass through. OpenAI has patched the bug and contributed the fix back to the Redis maintainers. The company also says it’s added extensive testing around this specific failure mode to prevent a repeat. For a service that gained 100 million users faster than any app in history, the breach was remarkably contained, but it’s a sharp reminder that caching layers in high-scale distributed systems are absolutely unforgiving of even subtle concurrency bugs.