If you install OpenAI’s ChatGPT Atlas, you WILL get hacked. Same applies to any AI browser. It’s not an if you will, it’s a when.
I wanted to make this clear on the very first line of this article to emphasize just how dangerous the very idea is. It has nothing to do with “AGI” and even for those who are lazy about privacy(which you shouldn’t be) should be concerned here, as this stupid idea can lead to bad actors stealing your bank account, e-mail, all your passwords, crypto, etc.
Mechanism
This is the way it’d work, as explained by a post on Brave’s blog:
- Initial Preparation: The attacker plants hidden directives within web pages using techniques like color camouflage (white text on a white background), HTML annotations, or other non-visible components. They might also infiltrate these directives into content created by users on platforms like Reddit or Facebook.
- Activation: A naive user visits the compromised web page and activates the AI assistant feature in their browser, perhaps by pressing a button to condense the page’s information or asking the AI to highlight important details.
- Command Insertion: During the AI’s analysis of the web page, it encounters these concealed commands. The AI fails to differentiate between the information it’s meant to summarize and the embedded instructions it shouldn’t execute, treating all as legitimate user requests.
- Malicious Execution: The AI follows the injected commands, which may involve misusing its browsing capabilities. For instance, it could be directed to access the user’s online banking, retrieve stored login credentials, or send sensitive data to a server managed by the attacker.
There’s also privacy
Another massive issue with this kind of browser is privacy. It doesn’t matter how much AI companies claim they won’t train, they already break the law left and right, specially with stolen content, they are guaranteed to train data on you.
And the data they may use to train could include sensitive information, on purpose or not, such as:
- Login/Passwords
- Real name
- Location
- Government issued documents
- Habits and Search History
- Voice
- Photos
- Videos
and the list goes on.
Once acquired, all models from that company will train on the data. All users of the model will have access to your data if they ask. There’s no way to avoid this entirely.
So, be it OpenAI, Anthropic, Meta, Google, Microsoft or whoever: if you see an AI focused browser, run. You will lose a lot more than gain by installing it.
Source: Brave
