Atlas Shrugged, the web flinched

Oh look. A new tech revolution: ChatGPT Atlas. So shiny. Much shareholders.

"With Atlas, ChatGPT can come with you anywhere across the web—helping you in the window right where you are, understanding what you’re trying to do, and completing tasks for you, all without copying and pasting or leaving the page."

Sam Altman must be really innovating. Let's take a look under the hood of these beast. What's that? Oh, it's built off Chromium? And they skipped the legal requirement for attribution in derivative work?

Promises.dmg

Gasp. Shock. Okay, what's next? How about some highlights from the .dmg installer screen:

1. Turn on browser memories Allow ChatGPT to remember useful details as you browse to give smarter responses and proactive suggestions. You're in control - memories stay private.

*:laughs in legally grey: The first known warrant for ChatGPT prompts was issued this week. *

2. Ask ChatGPT - on any website Open the ChatGPT sidebar on any website to summarize, explain, or handle tasks - right next to what you're browsing.

The average page is like… 500 words, guys. Come on.

3. Make your cursor a collaborator ChatGPT can help you draft emails, write reviews, or fill out forms. Highlight text inside a form field or doc and click the ChatGPT logo to get started.

Ah yes, the madlib technique. I look forward to reading FartBurrito69's gripping product reviews

4. Set as default browser BOOST CHATGPT LIMITS Unlock 7 days of extended limits on messaging, file uploads, data analysis, and image generation on ChatGPT Atlas.

There's that freemium carrot.

5. You're all set — welcome to Atlas! Have fun exploring the web with ChatGPT by your side, all while staying in control of your data and privacy.

This screen also displays shareable PNG badge with days since you registered for ChatGPT and Atlas. Because that somehow matters. Add it to your LinkedIn! If it's your profile picture, I'll know to ignore that message immediately. Thanks for saving me time, OAI!

Safety not Guaranteed

The landing page is more lofty promises than tech details. There is some information on how the browser is involved in training though:

"By default, we don’t use the content you browse to train our models. If you choose to opt-in this content, you can enable “include web browsing” in your data controls settings. Note, even if you opt into training, webpages that opt out of GPTBot, will not be trained on."

ChatGPT agent is native in Atlas (in preview to Plus, Pro and Business users. Terms and restrictions may apply.).That's the one that was released in July. You know, the one that took an hour to order food and recommending visiting a baseball stadium in the ocean. Apparently it's better now, but you still shouldn't trust it to do your grocery shopping.

The company seems quite aware of the risks involved with agentic AI, writing:

*"ChatGPT's agent capabilities still carry risk. **Besides simply making mistakes when acting on your behalf, agents are susceptible to hidden malicious instructions, which may be hidden in places such as a webpage or email with the intention that the instructions override ChatGPT agent’s intended behavior. *This could lead to stealing data from sites you're logged into or taking actions you didn't intend."

Here are some of the safety guardrails they've put in:

• It cannot run code in the browser, download files, or install extensions
• It cannot access other apps on your computer or file system
• It will pause to ensure you're watching it take actions on specific sensitive sites such as financial institutions
• You can use agent in logged out mode to limit its access to sensitive data and the risk of it taking actions as you on websites

Want to know the ultimate secret to optimizing your site for your site? "Website owners can also add ARIA ⁠(opens in a new window)tags to improve how ChatGPT agent works for their websites in Atlas."

A user-agent by any other name

K. Super fun. How does it behave though? Surely people with less trust issues than me have already downloaded it! Let's start with that user agent. Pedro Dias asked Atlas what his user agent was and got the response:

ChatGPTBrowser Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36

(lol. Chrome.)

This got awkward when he checked the user agent and got:

Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36

Why so shy, Atlas? You were all proud of your name just a second ago.

Then Tyler Gargula chimed in that Agent Mode mostly used his IP with a chrome user-agent, while sometimes using a ChatGPT-User agent under a different, yet consistent IP.

Jean-Christophe Chouinard observed that reverse DNS for Atlas requests came back as Googlebot. Oleg Korneitchouk chimed in:

*"I'm seeing mutiple user agents when I visit my homepage -- looks like Atlas is...

1. visiting the page entered via ChatGPT-User user agent (OAI IP) , then
2. sends Googlebot to the robots.txt (googlebot IP), then
3. hits the homepage again with a "normal" user agent (Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36) (local IP)
4. all page resources are loaded via 'normal' UA AND googlebot it really does seem like googlebot is visiting the site in parallel. testing on a new orphaned url next"*

Oleg then built an orphaned page on a domain with no robots.txt. This introduced the OAI-SearchBot crawler which repeatedly requested the robots.txt. Googlebot wasn't triggered.
He hypothesizes that:

1. OAI-SearchBot is used when you have 'web search' enabled in the browser. "Looks like it tries to 're-pull' the content remotely when this is enabled."
2. ChatGPT-User is used when you chat with the page. ChatGPT uses screenshots of pages for its analysis. This UA doesn't execute javascript.

He also got a new user agent:

ChatGPT%20Atlas/20251021184832000 CFNetwork/3860.100.1 Darwin/25.0.0

"I've seen 0 GPTBot activity thus far."

And then Atlas warned him it detected 'unusual activity from my device'. What do all these user agents mean for the promise made with "Note, even if you opt into training, webpages that opt out of GPTBot, will not be trained on."?

What user-agent does ChatGPT Atlas use?

Here are the ChatGPT Atlas user-agents we know about (so far):

1. Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36

2. ChatGPTBrowser Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36

3. ChatGPT%20Atlas/20251021184832000 CFNetwork/3860.100.1 Darwin/25.0.0

JC is activity updating the post as he learns more. St. Oleg, thank you for your service. I owe you one.

OpenAI you owe me– and every one else in digital— technical documentation and a damn apology.

The broken bot covenant

Youngins, gather round. We need to divert for a second back to the once-sacred ruling law of the internet: the bot covenant. It's always cost sites money to handle crawlers and scrapers have been stealing content since the first geocities blossomed. But we had an agreement. We would allow scrapers like search engines to access our content and in exchange, the bot would do 3 things:

1. Crawl politely
2. Declare who they were
3. Send you traffic in return

Crawl politely? Crawl-to-refer ratio for OpenAI is about 800:1. In 24 hours, we clearly established that this browser is not going to properly declare itself. Send you traffic? See 800:1 above.

This is nonsense. This is cobbled together technology powered by the will of shareholders. It's the equivalent of a Cybertruck with a Rivian badge– badge doesn't matter when someone's going to lose a finger.

Agentic AI has a laundry list of search vulnerabilities. It's never met a phishing link it didn't want to click. Hell, it might blackmail you.

Accountability not found, please try again later

Who gets held accountable for their learn-fast-break-things mistakes? These companies are straddling the line between publisher and platform. Unchecked, these platforms are set for unprecedented access and impact without quantifiable accountability. Publishers are legally liable for the content they publish, including defamation or copyright infringement. Platforms are generally not liable for user-generated content, thanks to legal protections like Section 230.

It doesn't have to be this way and it's not in other countries. I'd like to end this with a speech from German Federal Government Commissioner for Culture and the Media, Wolfram Weimar (h/t Goran Majić for the translation).

"In a manner akin to vampirism, AI companies are currently sucking the creative potential out of countless brilliant minds, exploiting their ideas and emotions, their creative power, their visions. This turns the great cultural achievement of autonomous artworks—and above all, books—into prey. I use the word deliberately—I consider this intellectual vampirism. And it strikes literature.

Authors who have wrestled over every single line, whose entire existence sometimes revolved around the written word, who have poured themselves into it, their life's work, their creative energy, even their crises and doubts—all of that is now turned into a vast pool from which anything needed is fished out at will to generate huge revenues, sometimes revenues in the billions.

But even the term data mining strikes me as a euphemism. For it suggests that the big Big Techs, that the AI companies, are somehow digital descendants of mining companies, as if they had a license to extract raw materials and process them further. But what we're dealing with here is not raw materials—it’s texts, images, music, films, entire life's works. And above all: they don’t have a license.

There’s a gold rush atmosphere prevailing. But what’s happening in the data centers from Silicon Valley to Shenzhen is a highly lucrative, industrially organized, technologically super impressive—but ultimately—a raid. Let me say it plainly: The American and Chinese tech giants are training their AI systems with the contents and creative achievements of this world. But they didn’t ask us. And that’s why entire cultures are now being degraded to raw material suppliers and actually shamelessly exploited.

I call this digital colonialism, which we should no longer tolerate.”

We can do better.