Last year, OpenAI introduced a search feature into ChatGPT, allowing users to find information. Fast forward, the idea evolved,
and now last week they surprised the world by introducing the first fully fledged AI-powered browser. It works like Google’s
Chrome or Apple’s Safari. For a while, Google has integrated Chrome with its AI LLM Gemini, which appears on top of your
Google search. Currently, 60% of internet users use Google Chrome, and OpenAI is looking into tapping into that market, but it
won’t be an easy task.
They describe Atlas as a super assistant that can learn your world and consequently assist you in navigating through it easily. It
is built in ChatGPT, so you don’t need to open or navigate different tabs or copy and paste your searches. The selling point is
that as you continue to use Atlas, it becomes smarter by memorising context so that whenever you browse something, you can
see the most relevant depending on your interest and search history. For example. If you are searching for a particular website,
Atlas can filter the content of the site to suggest the article that may be of interest to you. If you search for a particular recipe,
for example, Atlas can contextualize that and recommend the nearest grocery store where you can find all the ingredients that
you will need.
Atlas Browser can work using agentic mode as well. This helps to improve speed and information processing. Currently, the
new browser will be available to macOS devices, including the free, plus, and go users.
Most AI companies are still struggling to raise revenue, and for sure, this is another attempt to create a new revenue stream for
OpenAI. To make the most use of Atlas features, you will need to pay for it. It remains to be seen how many users will be willing
to part with their dollars for a browser, as only 5% of ChatGPT users who are currently on a paid subscription version.
Despite all the glam and convincing from OpenAI after the launch, the security community has raised concerns about the bar
address, which can be used for both URLs and prompts . This invokes new security risks, especially in relation to prompt
injection attacks where malicious instructions camouflaged as a link could trick the AI into performing unintended actions.
Additionally, despite OpenAI stating that they won’t be using information from Atlas for its AI model training, it remains to be
seen how they are going to navigate through this. No matter how you look at it, for those who treasure their data and
anonymity, this might be too big a risk to take
