RC RANDOM CHAOS

OpenAI's __obi cookie links your off-site browsing to your ChatGPT account

· via Hacker News

Original source

ChatGPT now knows what you do on other websites via ad collector

Hacker News →

A researcher has documented how OpenAI’s advertising infrastructure quietly connects activity on third-party websites back to individual ChatGPT users. When you use ChatGPT, the client mints a short-lived signed JWT that binds your account (or a stable anonymous device ID) to a random identifier called ‘obi’, then exchanges it at bzr.openai.com — OpenAI’s internal ‘bazaar’ ad platform — for a cookie named __obi. That cookie is scoped to .openai.com with SameSite=None, the one configuration that lets it ride along on cross-site requests, and it lasts a year. Any site running OpenAI’s advertising pixel then leaks __obi back to OpenAI on page load, along with what you searched, read, or bought. Notably, the identifier is disclosed the moment the SDK script tag loads, before any ‘credential-omitting’ code path can run.

The pixel does more than track conversions. Its SDK scrapes identity directly from advertiser pages — pulling emails and phone numbers from the tag-manager dataLayer and form fields, and hashing email, phone, and names with SHA-256 while sending location data in the clear. In observed traffic, scraped identity outnumbered advertiser-supplied identity nearly three to one, and URL paths reaching the collector exposed sensitive contexts including a medical condition, a debt-relief funnel, and a litigation intake form. The researcher reproduced the mechanism across roughly a thousand pixels and confirmed a single __obi value being sent from a dozen well-known retailers such as Chewy, Wayfair, HelloFresh, and Coursera. Automatic identity matching was on for most pixels, including every credit and lending advertiser observed.

Two details sharpen the privacy concern. First, __obi is classified in OpenAI’s cookie policy as an ‘analytics’ cookie, and every sync token carried consent_decision: analytics_allowed — meaning users who granted analytics but explicitly refused marketing still received it. OpenAI Support acknowledged the researcher’s inquiry but declined to answer why the cookie is labeled analytics or whether refusing marketing stops it. Second, while the underlying adtech (a logged-in account resolved to off-site behavior via third-party cookies) is structurally identical to what Meta and Google have run for years, applying it to an AI chat product is unprecedented. People disclose things to ChatGPT they would never post to a social network, and the assistant increasingly acts on their behalf — making this join far more intimate than conventional ad tracking. The mechanism was observed only on Chrome for Android; Safari’s tracking prevention and iOS WebKit block it, and only about one in five sessions produced a sync token.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.