A library bug briefly mixed up ChatGPT users' account data

Start here

An outage report is a rare chance to see, in a vendor's own words, what a chat service actually stores about an account, because the description of what went wrong doubles as a description of what is normally there to break. OpenAI's account of a March 2023 bug is useful for that reason, beyond the incident itself.

What the documents say

OpenAI's own report, published 24 March 2023, states that a bug in an open-source library, redis-py, caused some users to see chat titles from another active user's history, and possibly the first message of a newly created conversation. The company explains the mechanism plainly: a shared pool of connections to its Redis cache could become corrupted when a request was cancelled at a specific moment, so 'the next response that's dequeued for an unrelated request can receive data left behind in the connection.' The report discloses a second, more serious effect: for roughly 1.2% of ChatGPT Plus subscribers active during a specific nine-hour window, another active user's name, email address, payment address, card type, last four card digits and expiration date could have briefly appeared, though 'full credit card numbers were not exposed at any time.' OpenAI's Data Controls FAQ, as retrieved on 16 September 2026, describes the ordinary version of the system this bug disrupted: chat history and account settings that persist on OpenAI's servers by default.

Check this

The useful check here is a mechanism, not a number: a chat service typically holds session and billing information in a shared server-side cache to avoid repeatedly querying its main database, and a concurrency bug in the code managing that cache is what let one person's cached data leak into another person's response. That is a structural property of many web services, not a flaw unique to one company. A reader evaluating any AI provider's security claims can ask whether an incident history exists and whether the provider explains the specific mechanism, as this report does, rather than offering only a vague apology.

What holds and what fails

What holds is the value of a detailed postmortem: naming the library, the failure condition and the affected population lets outside engineers judge the claim, rather than take it on faith. What the report does not establish is a complete account of every person affected, since OpenAI states its confidence that exposure was 'extremely low' based on the specific steps required to see another user's data, a probabilistic judgment, not a full audit. Treat the company's own severity assessment as its claim, stated with a described basis, rather than as independently verified.

• Read a vendor's incident postmortems, if any exist, before trusting its security claims.
• Note that stored billing details, not just chat text, can be exposed by an application bug.
• Check whether a service publishes what data classes it holds in ordinary operation.

A single bug report will not tell you whether a company is secure in general. It will tell you, concretely, what categories of your data sit on its servers waiting to be exposed the next time something breaks.