Post

Wrote up my first impressions of ChatGPT Atlas, OpenAI's new browser - I remain unconvinced by the entire category of "browser agents", the security and privacy challenges still feel insurmountable to me simonwillison.net/2025/Oct/21/...

simonwillison.netIntroducing ChatGPT AtlasLast year OpenAI hired Chrome engineer Darin Fisher, which sparked speculation they might have their own browser in the pipeline. Today it arrived. ChatGPT Atlas is a Mac-only web browser …

18 reposts125 likes

Simon, wonder if you've had a chance to look at this yet, SDD kit from github bsky.app/profile/rami...

Rami Krispin@ramikrispin.bsky.social

GitHub Spec-Kit 🚀 GitHub recently released a new project - Spec-Kit, an open-source project that aims to supplement AI code generation tools such as Claude Code, OpenAI Codex, etc. 🔗: github.com/github/spec-... #ai #github #opensource

ChatGPT Atlas notwithstanding, do you find yourself using these tools like Claude Code Web, Codex CLI etc. more than your llm utility? How would you say your usage of llm has changed over time?

As of a couple of months I'm using Codex CLI and Claude Code CLI WAY more often than my own LLM utility - they've become my daily drivers for all sorts of things Partly that's because of cost: Codex and Claude Code both bill against my existing subscriptions, no extra API charges

You worry more about prompt injection than the most important security (and privacy) issue: data collection. The amount and quality of information in prompts is far bigger than any social network or Google receives. Or does the Atlas run models locally? Each prompt is a potential confession.

I'm just personally not that worried about the data collection side of things Prompt injection allows a malicious criminal to directly target me for harm - I find that a lot more concerning than OpenAI and co using my personal data for creepy personalized ads

It's not just about ads. It's more about what a dictatorial/fascist government could force OpenAI to do with your data. Remember that LLMs can analyse prompts for all kinds of things as they come in. You would not know about it until you are denied something or someone knocks at your door.

I do get concerned when I think about the journalism side of this - journalists need to protect their sources, and a hostile government could absolutely show up at a data center with a warrant for information relating to that

Meanwhile, a lot of us, including me, believe that genAI-assisted accessibility is a promising way forward, and we're waiting for not-BS tools to get into production. This extended use of aria might be a good step.

I'm stuck on who it's for. Browsers are easy to use and most tasks done with them are already fast. If you want automatic stuff we do it with APIs much cheaper. Examples I've seen take more user time than it would have taken to just do it! What's the point apart from showing you can?

I suspect the 'showing you can' is a big part of it tbh (the browser agent part). I mean it seems to be a very neat trick without a very good use case. Security-wise it just seems like a recipe for disaster. I don't see how it could be locked down enough to let er rip in anything persistent

I work on websites and we make them for humans - if we only have to make them for computers to use it would be so much easier! I also just can't fathom why we would use a computer to access the HCI which is the hard part - making it hard for ourselves from both directions!