Mixpanel was breached. OpenAI API users are in the phishing blast radius.

OpenAI confirmed that a breach at analytics vendor Mixpanel exposed metadata about some users of its API. This is not a story about a model escaping a sandbox. It is a story about a third party that should never have been the most interesting sentence in an AI company’s week.

What was taken

Mixpanel said an unknown attacker got into part of its systems around 8 November and exported a dataset. OpenAI’s account of what that dataset held for API customers: names, email addresses, approximate browser-based location, operating system, and browser details.

OpenAI says the export did not include prompts, API keys, payment information, or authentication tokens. ChatGPT users who only use the consumer chatbot in the browser are outside this incident. The exposed population is people who hit OpenAI through the API: the developers and the companies built on top of the API.

OpenAI is dropping Mixpanel.

Why “only metadata” still matters

Attackers do not need your system prompt to make this useful. They need a list of people who can be convinced that OpenAI, or a “security review of your API account,” needs a token, a password, or a one-time code.

Name plus email plus “uses OpenAI API” is a targeting packet. Expect credential-harvesting mail within days, not months.

If you are an API customer, rotate anything that could be phished, warn the humans who own those inboxes, and look at which other vendors still receive identifiable customer fields they do not need.

The AI-specific lesson is small and unfashionable

Frontier labs can publish model cards and still fail the oldest exam in vendor risk: which subprocessors see identity data, and what happens when that subprocessor is the one that gets hit.

Keep this incident in the “supply chain” folder, not the “rogue agent” folder. Mixing the two makes both problems harder to brief to a board.