Understanding the Risks of AI Browsers

AI browsers have revolutionized the way we interact with the web, making it easier to perform complex tasks and execute actions autonomously. However, these browsers also expose users to new and unique security risks. One of the most significant threats is prompt injection, where malicious AI prompts are hidden in a website or browser extension’s HTML source code.

Source: lifehacker.com

When a user navigates to a web page with these hidden prompts, their browser will read the instructions and automatically execute the actions without asking for verification or consent. This can lead to serious consequences, such as accessing sensitive information, logging in to unauthorized accounts, or even completing financial transactions without the user’s knowledge.

Source: lifehacker.com

The Risks of AI Prompt Injection

AI browsers like Perplexity Comet and ChatGPT Atlas are particularly vulnerable to prompt injection attacks. In fact, the official OpenAI documentation warns against using Atlas with production data due to the risk of prompt injection. The Brave security team has also demonstrated the risks of CometJacking, where the browser can be manipulated into accessing user email addresses, obtaining one-time passwords, and forwarding them to attackers.

Another significant risk is cross-site request forgery (CSRF), where a malicious web page can send instructions to the browser as if the user had typed them themselves. This can lead to serious consequences, including accessing sensitive information, logging in to unauthorized accounts, or even completing financial transactions without the user’s knowledge.

Enabling Built-in Browser Settings for Better Safety

While AI browsers carry many vulnerabilities and loopholes, there are legitimate ways to experiment with them without compromising your privacy. Most of these browsers have optional features that can be enabled to beef up security and prevent unauthorized access to your data.

One of the most effective ways to protect yourself is to disable data sharing. Almost every AI browser uses your browsing patterns and search history to train future iterations of its AI models. This means that all your browsing data is being sent to the browser’s developers by default unless you specifically opt out. Luckily, browsers that train models on your data also give you the option to disable training, at least on paid plans.

For example, ChatGPT Atlas allows you to disable model training by navigating to Settings > Data Controls and turning off Improve model for everyone. Perplexity Comet also offers this option by going to New Tab Page > Account > Preferences and toggling off AI data retention.

Preventing Unauthorized Access to Your Logged-in Sessions

Another significant risk is that AI browsers can be manipulated into accessing your logged-in accounts on different websites and retrieving sensitive information through prompt injection. Depending on their level of access, they can also go into your accounts to execute certain actions without your knowledge, like sending an email or downloading a file.

ChatGPT Atlas offers a feature that allows you to specifically prevent the AI from accessing your logged-in browser sessions in Agent Mode. By choosing Agent mode and switching to Logged out, you can prevent the AI from accessing your active sessions by default, instead prompting you to log in manually if your task requires access to a user account.

Perplexity Comet does not have a universal toggle that restricts access to logged-in sessions. However, you can use Incognito mode when logging into any websites with the Comet browser to prevent staying logged in after you quit.

Conclusion

While AI browsers carry many vulnerabilities and loopholes, there are legitimate ways to experiment with them without compromising your privacy. By enabling built-in browser settings and taking steps to prevent unauthorized access to your logged-in sessions, you can safely use AI browsers and enjoy the benefits they offer.