ChatGPT's ad cookie wanders, Anthropic eyes a new model, Grok logs in.

On September 20, an independent security researcher who publishes as Buchodi posted a teardown of the tag behind ChatGPT ads. The claim: when a business installs OpenAI's measurement tag, a cookie called __obi lives for a year and is set so it can travel across sites. By the researcher's count, the tag collected identity data 685 times by scraping it from forms and pages, against 255 times when the advertiser supplied it on purpose. The data included hashed emails and phone numbers, plus unhashed city and postal code.

The part that will make a privacy lawyer put down their coffee: the tokens the researcher decoded were marked as analytics, even when the tester refused marketing cookies. The researcher says OpenAI acknowledged two questions on September 14 and has not answered them. Now the caveats, because shrewd beats loud. This is one researcher, not an OpenAI statement. iPhones were not affected, Android Chrome was, and desktop Chrome was not tested. Treat it as a strong lead, not a verdict.

Also on September 20, Reuters reported that Anthropic is considering releasing a new model, citing three sources, and Anthropic declined to comment. The backdrop is a scoreboard nobody at Anthropic enjoys: spending data from Ramp puts OpenAI's GPT-6 Astra at about 13% of tracked enterprise AI spend, against about 8% for Claude Fable. On OpenRouter, OpenAI models outspent Anthropic's for the first time in more than two and a half years.

It is not a rout. Reuters puts Anthropic's annualized revenue at $65 billion at the end of July, ahead of OpenAI's $40 billion. The same report says Anthropic's IPO could slip past the November elections. Translation: the two biggest labs are trading punches every few weeks, and your tools will keep changing underneath you.

Marketing insight: Stop marrying one model. Keep your brand voice guide, your best prompts and your customer facts in a plain document you own, so switching tools takes an afternoon instead of a rebuild.

SpaceXAI's Grok Bot is now in early beta, and Engadget's September 20 walkthrough spells out what it does. This is not the chatbot. It is an agent that signs into tools like Gmail, LinkedIn and Salesforce and acts for you: filling out forms, replying on social media, sending email. Engadget lists it for SuperGrok Plus and Heavy subscribers and for Cursor Pro+, Ultra and Teams plans, on macOS, Windows, iOS and Android.

Engadget's advice is refreshingly modest: keep human approval switched on for anything that could send a message, change a record or expose sensitive information, and start with small, low-risk jobs you supervise. In other words, the robot intern is here, and even the people cheering for it say do not hand it the company card on day one.

Put the three stories side by side and the pattern is hard to miss. AI is moving into your ads, your software choices and your inbox, all at once, and each move asks the same question: can your customer trust what happens on your behalf? The cookie story is about tags you did not read. The Reuters story is about vendors trading the lead every few weeks. The Grok story is about software that speaks in your name. The businesses that win this stretch will not be the ones with the most AI. They will be the ones who can explain, in one plain paragraph, what their AI does and does not do.

If you would like a second pair of eyes on your tags, your tools or your agent rules, our founders do this every week. Talk to them below. See you in the next edition.